@fugood/buttress-server 2.23.0-beta.37

package/README.md

@@ -0,0 +1,193 @@
+# Buttress Server
+
+A high-performance RPC server for managing GGML LLM generators with configurable defaults and runtime management.
+
+## Installation
+
+```bash
+npm install -g @fugood/buttress-server
+```
+
+## Quick Start
+
+### Using CLI
+
+```bash
+# Start with config file
+npx bricks-buttress --config ./config.toml
+
+# Start without config (uses env vars and defaults)
+npx bricks-buttress
+```
+
+## Configuration
+
+Configuration can be provided via:
+- `--config` / `-c` flag with TOML file path
+
+### Configuration Format (TOML)
+
+```toml
+# Environment variables (only set if not already defined in system)
+[env]
+HF_TOKEN = "your_huggingface_token_here"
+CUDA_VISIBLE_DEVICES = "0"
+
+[server]
+port = 2080
+log_level = "info"
+
+[runtime]
+cache_dir = "~/.buttress/models"
+
+# Session state cache for ggml-llm (saves KV cache to disk for prompt reuse)
+[runtime.session_cache]
+enabled = true
+max_size_bytes = "10GB"  # Supports string (e.g., "10GB", "500MB") or number
+max_entries = 1000
+
+# GGML LLM generator
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant_preference = ["cuda", "vulkan", "default"]
+[generators.model]
+repo_id = "ggml-org/gpt-oss-20b-GGUF"
+quantization = "mxfp4"
+n_ctx = 12800
+
+# GGML STT (Speech-to-Text) generator
+[[generators]]
+type = "ggml-stt"
+[generators.backend]
+variant_preference = ["coreml", "default"]
+[generators.model]
+repo_id = "BricksDisplay/whisper-ggml"
+filename = "ggml-small.bin"
+```
+
+### Programmatic Usage
+
+```javascript
+import { startServer } from '@fugood/buttress-server'
+
+startServer({
+  port: 3000,
+  defaultConfig: {
+    runtime: {
+      cache_dir: './.buttress-cache'
+    },
+    generators: [
+      {
+        type: 'ggml-llm',
+        model: {
+          repo_id: 'ggml-org/gemma-3-270m-qat-GGUF',
+          quantization: 'mxfp4',
+        }
+      }
+    ]
+  }
+})
+  .then(({ port }) => {
+    console.log(`Server running on port ${port}`)
+  })
+  .catch(console.error)
+```
+
+### Environment Variable Priority
+
+Environment variables can be set in the `[env]` section of the TOML config. These values will **only be applied if the environment variable is not already set** in the system. This allows:
+
+1. **Default values** in config file
+2. **System environment variables** to override config values
+3. **Command-line exports** to have highest priority
+
+Example:
+```bash
+# Config has: [env] HF_TOKEN = "default_token"
+
+# This will use the system env variable (highest priority)
+HF_TOKEN=my_token npx bricks-buttress
+
+# This will use the config value
+npx bricks-buttress
+```
+
+## Port Priority
+
+Port can be configured via multiple sources (highest priority first):
+
+1. **Command-line flag**: `--port 3000`
+2. **Config file**: `[server] port = 2080`
+3. **Default**: `2080`
+
+## CLI Reference
+
+```
+bricks-buttress v2.23.0-beta.22
+
+Buttress server for remote inference with GGML backends.
+
+Usage:
+  bricks-buttress [options]
+
+Options:
+  -h, --help                    Show this help message
+  -v, --version                 Show version number
+  -p, --port <port>             Port to listen on (default: 2080)
+  -c, --config <path|toml>      Path to TOML config file or inline TOML string
+
+Testing Options:
+  --test-caps <backend>         Test model capabilities (ggml-llm or ggml-stt)
+  --test-caps-model-id <id>     Model ID to test (used with --test-caps)
+  --test-models <ids>           Comma-separated list of model IDs to test
+  --test-models-default         Test default set of models
+
+  Note: --test-models and --test-models-default output a markdown report
+        file (e.g., ggml-llm-model-capabilities-YYYY-MM-DD.md)
+
+Environment Variables:
+  BUTTRESS_PORT                 Port to listen on (overridden by --port)
+  NODE_ENV                      Set to 'development' for dev mode
+
+Examples:
+  bricks-buttress
+  bricks-buttress --port 3000
+  bricks-buttress --config ./config.toml
+  bricks-buttress --test-caps ggml-llm --test-models-default
+  bricks-buttress --test-caps ggml-stt --test-caps-model-id BricksDisplay/whisper-ggml:ggml-small.bin
+```
+
+## Session State Cache
+
+The server supports session state caching for ggml-llm generators, which saves KV cache state to disk after completions. This enables:
+
+- **Prompt reuse**: Same or similar prompts can reuse cached state, skipping prompt processing
+- **Multi-turn conversations**: Conversation history state is preserved across requests
+
+### Configuration
+
+```toml
+[runtime.session_cache]
+enabled = true                  # Enable/disable session caching (default: true)
+max_size_bytes = "10GB"         # Supports string (e.g., "10GB", "500MB") or number (default: 10GB)
+max_entries = 1000              # Max number of cached entries (default: 1000)
+```
+
+### How it works
+
+1. After a successful completion, the KV cache state is saved to disk
+2. On new completions, the server checks if any cached state matches the prompt prefix
+3. If a match is found, the cached state is loaded, skipping redundant prompt processing
+4. LRU eviction removes oldest entries when limits are exceeded
+
+### Cache location
+
+Cache files are stored in `{cache_dir}/.session-state-cache/`:
+- `cache-map.json` - Index of cached entries
+- `states/` - Binary state files
+- `temp/` - Temporary files (auto-cleaned after 1 hour)
+
+## Tips
+
+- macOS: Use `sudo sysctl iogpu.wired_limit_mb=<number>` to increase GPU memory allocation. The default available memory of GPU is about ~70%. For example, if the hardware have 128GB memory, you can use `sudo sysctl iogpu.wired_limit_mb=137438` to increase to 128GB. Run `sudo sysctl iogpu.wired_limit_mb=0` if you want to back to default.

package/config/sample-iq9-cpu.toml

@@ -0,0 +1,93 @@
+# Sample Buttress server configuration with per-model defaults
+
+# Environment variables (only set if not already defined)
+[env]
+# HF_TOKEN = "your_huggingface_token_here"
+# CUDA_VISIBLE_DEVICES = "0"
+
+[server]
+port = 2080
+log_level = "info"
+# max_body_size = "100MB"  # Supports string (e.g., "100MB", "1GB") or number in bytes
+
+[runtime]
+cache_dir = "./.buttress-cache"
+# huggingface_token = "hf_xx"
+
+# Global model params
+n_batch = 512
+n_ubatch = 512
+flash_attn_type = "off"
+cache_type_k = "f16"
+cache_type_v = "f16"
+n_threads = 8
+
+# Session state cache for ggml-llm (saves KV cache to disk for prompt reuse)
+[runtime.session_cache]
+enabled = true
+max_size_bytes = "20GB"  # Supports string (e.g., "10GB", "500MB") or number
+max_entries = 1000
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant = "default"
+variant_preference = ["default"]
+gpu_memory_fraction = 0.8
+cpu_memory_fraction = 0.8
+[generators.model]
+repo_id = "ggml-org/gpt-oss-20b-GGUF"
+quantization = "mxfp4"
+download = true
+n_ctx = 51200
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant = "default"
+variant_preference = ["default"]
+gpu_memory_fraction = 0.8
+cpu_memory_fraction = 0.8
+[generators.model]
+repo_id = "unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF"
+quantization = "q4_0"
+download = true
+n_ctx = 51200
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant = "default"
+variant_preference = ["default"]
+gpu_memory_fraction = 0.8
+cpu_memory_fraction = 0.8
+[generators.model]
+repo_id = "unsloth/Qwen3-4B-GGUF"
+quantization = "q4_0"
+download = true
+n_ctx = 51200
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant = "default"
+variant_preference = ["default"]
+gpu_memory_fraction = 0.8
+cpu_memory_fraction = 0.8
+[generators.model]
+repo_id = "ggml-org/gemma-3-12b-it-qat-GGUF"
+quantization = "q4_0"
+download = true
+n_ctx = 51200
+
+# Speech-to-Text (STT) generators
+[[generators]]
+type = "ggml-stt"
+[generators.backend]
+variant = "default"
+variant_preference = ["default"]
+[generators.model]
+repo_id = "BricksDisplay/whisper-ggml"
+filename = "ggml-base.bin"
+download = true
+use_gpu = false

package/config/sample-opencode.json

@@ -0,0 +1,17 @@
+{
+  "$schema": "https://opencode.ai/config.json",
+  "provider": {
+    "buttress": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Buttress (GPT-OSS 20B)",
+      "options": {
+        "baseURL": "http://localhost:2080/oai-compat/v1"
+      },
+      "models": {
+        "ggml-org/gpt-oss-20b-GGUF": {
+          "name": "gpt-oss-20b"
+        }
+      }
+    }
+  }
+}

package/config/sample.toml

@@ -0,0 +1,138 @@
+# Sample Buttress server configuration with per-model defaults
+#
+# Model Pre-Download:
+#   Set `download = true` on a generator's [generators.model] section to
+#   automatically download the model at server startup if it doesn't exist.
+#   When initContext is called, it will wait for any in-progress download
+#   of the same model before proceeding.
+
+# Environment variables (only set if not already defined)
+[env]
+# HF_TOKEN = "your_huggingface_token_here"
+# CUDA_VISIBLE_DEVICES = "0"
+
+[autodiscover.http]
+enabled = true
+
+[server]
+port = 2080
+log_level = "info"
+# max_body_size = "100MB"  # Supports string (e.g., "100MB", "1GB") or number in bytes
+
+[openai_compat]
+# cors_allowed_origins = ["http://localhost:3000", "https://example.com"]  # Restrict to specific origins
+# cors_allowed_origins = "*"  # Allow all origins (default)
+
+[runtime]
+cache_dir = "./.buttress-cache"
+# huggingface_token = "hf_xx"
+
+# Global model params
+flash_attn_type = "on"
+cache_type_k = "q8_0"
+cache_type_v = "q8_0"
+n_threads = 6
+
+# Delay before releasing context when refCount reaches 0 (default: 10000ms)
+# Set to 0 for immediate release
+context_release_delay_ms = 30000
+
+# Session state cache for ggml-llm (saves KV cache to disk for prompt reuse)
+[runtime.session_cache]
+enabled = true
+max_size_bytes = "10GB"  # Supports string (e.g., "10GB", "500MB") or number
+max_entries = 1000
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant_preference = ["cuda", "vulkan", "snapdragon", "default"]
+gpu_memory_fraction = 0.95
+cpu_memory_fraction = 0.95
+[generators.model]
+repo_id = "ggml-org/gpt-oss-20b-GGUF"
+quantization = "mxfp4"
+download = true
+# n_ctx = 12800 # Max: 131072
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant_preference = ["cuda", "vulkan", "snapdragon", "default"]
+gpu_memory_fraction = 0.95
+[generators.model]
+repo_id = "ggml-org/gpt-oss-120b-GGUF"
+quantization = "mxfp4"
+download = false
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant_preference = ["cuda", "vulkan", "snapdragon", "default"]
+gpu_memory_fraction = 0.95
+[generators.model]
+repo_id = "unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF"
+quantization = "q4_0"
+download = true
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant = "default"
+variant_preference = ["cuda", "vulkan", "snapdragon", "default"]
+gpu_memory_fraction = 0.8
+cpu_memory_fraction = 0.8
+[generators.model]
+repo_id = "unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF"
+quantization = "q4_0"
+n_ctx = 51200
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant_preference = ["default"]
+gpu_memory_fraction = 0.95
+[generators.model]
+repo_id = "bartowski/mistralai_Devstral-2-123B-Instruct-2512-GGUF"
+quantization = "q4_0"
+download = false
+# n_ctx = 128000 # Max: 262144
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant_preference = ["default"]
+[generators.model]
+repo_id = "ggml-org/gemma-3-270m-qat-GGUF"
+quantization = "q4_0"
+download = true
+
+[[generators]]
+type = "ggml-llm"
+[generators.backend]
+variant_preference = ["default"]
+[generators.model]
+repo_id = "ggml-org/gemma-3-12b-it-qat-GGUF"
+download = false
+
+# Speech-to-Text (STT) generators
+[[generators]]
+type = "ggml-stt"
+[generators.backend]
+variant_preference = ["cuda", "vulkan", "default"]
+[generators.model]
+repo_id = "BricksDisplay/whisper-ggml"
+filename = "ggml-small-q8_0.bin"
+download = true
+use_gpu = true
+
+[[generators]]
+type = "ggml-stt"
+[generators.backend]
+variant_preference = ["cuda", "vulkan", "default"]
+[generators.model]
+repo_id = "BricksDisplay/whisper-ggml"
+filename = "ggml-large-v3-turbo-q8_0.bin"
+download = false
+use_gpu = true
+use_flash_attn = true

package/lib/index.d.ts

@@ -0,0 +1,27 @@
+import type { AnyElysia } from "elysia";
+import * as backendCore from "@fugood/buttress-backend-core";
+import { AutodiscoverService } from "./autodiscover";
+import type { Config } from "./types";
+export { startModelDownload } from "@fugood/buttress-backend-core";
+export { processConfig } from "./utils/config";
+export declare const checkForUpdates: () => Promise<string | null>;
+export declare const compareVersions: (current: string, latest: string) => boolean;
+export declare const logUpdateMessage: (latestVersion: string) => void;
+export declare const checkAndNotifyUpdates: () => Promise<void>;
+export type Backend = typeof backendCore;
+export interface StartServerOptions {
+	backend?: Backend;
+	router?: AnyElysia;
+	config: Config;
+	enableOpenAICompat?: boolean;
+}
+export declare const createServer: ({ backend, router, config, enableOpenAICompat }: StartServerOptions) => Promise<{
+	app: AnyElysia;
+	config: Config;
+}>;
+export declare const startServer: ({ backend, router, config, enableOpenAICompat }: StartServerOptions) => Promise<{
+	app: AnyElysia;
+	port: number;
+	openaiEnabled: boolean;
+	autoDiscover: AutodiscoverService | null;
+}>;

package/lib/index.mjs

@@ -0,0 +1,110 @@
+import{createRequire as x3}from"node:module";var M3=Object.defineProperty;var F3=($,Z)=>{for(var X in Z)M3($,X,{get:Z[X],enumerable:!0,configurable:!0,set:(j)=>Z[X]=()=>j})};var S=($,Z)=>()=>($&&(Z=$($=0)),Z);var Q4=x3(import.meta.url);var J4=($,Z,X)=>Math.min(Math.max($,Z),X),N4=($)=>$?40:0,H4=($=0)=>{if(!$)return 0;return J4($/12884901888*20,0,20)},V4=($=0)=>{if(!$)return 0;return J4($/34359738368*10,0,10)},U4=($)=>$?10:0,G4=($="default",Z=null)=>{let X=String($).toLowerCase();if(!X)return 0;if(X.includes("cuda"))return 20;if(X.includes("vulkan"))return 10;if(X.includes("default"))return Z==="darwin"||Z==="ios"?15:5;return 0},K0=({platform:$,variant:Z,hasGpu:X,gpuUsableBytes:j=0,cpuUsableBytes:W=0,ok:H=!0}={})=>{if(!H)return 0;let Q=N4(X)+G4(Z,$)+H4(j),J=V4(W),N=U4(H);return Math.min(100,Math.round(Q+J+N))},K1=({platform:$,variant:Z,hasGpu:X,gpuUsableBytes:j=0,cpuUsableBytes:W=0,ok:H=!0}={})=>({gpuPresence:N4(X),variant:G4(Z,$),gpuMemory:H4(j),cpuMemory:V4(W),availability:U4(H)});var O4,D0=0.85,b0=0.5,K4=($)=>{if(!$&&$!==0)return[];if(Array.isArray($))return $.filter((Z)=>Z!=null);return[$]},S3=($)=>{if(!$)return null;return String($).trim().toLowerCase()||null},P3=({variant:$,preferVariants:Z=[],variantPreference:X=[],defaultVariants:j=O4}={})=>{let W=[];if($)W.push($);W.push(...K4(Z)),W.push(...K4(X)),W.push(...j);let H=W.map(S3).filter(Boolean);return Array.from(new Set(H))},_4=($={})=>{let Z=String($.type||$.deviceType||$.kind||"").toLowerCase();if(Z.includes("gpu"))return!0;if(Z.includes("cuda"))return!0;if(Z.includes("metal"))return!0;if(Z.includes("vulkan"))return!0;if(Z.includes("snapdragon"))return!0;return!1},T3=($)=>{if(!Array.isArray($))return[];return $.map((Z)=>({...Z}))},k3=($,Z)=>{if($==="snapdragon")return Z.filter((X)=>X.deviceName!=="GPUOpenCL");return Z},q4=({platform:$,totalMemoryInBytes:Z,variant:X,devices:j,gpuMemoryFraction:W,cpuMemoryFraction:H,ok:Q,error:J})=>{let N=T3(k3(X,j)),G=N.some(_4),_=N.filter((w)=>_4(w)&&Number.isFinite(Number(w.maxMemorySize))).reduce((w,U)=>w+U.maxMemorySize,0),O=Z,z=G?Math.floor(_*W):0,V=O?Math.floor(O*H):0,R={platform:$,variant:X,hasGpu:G,gpuUsableBytes:z,cpuUsableBytes:V,ok:Q},q=K0(R),A=Q?K1(R):null;return{platform:$,ok:Q,variant:X,hasGpu:G,devices:N,gpuTotalBytes:_,gpuUsableBytes:z,cpuTotalBytes:O,cpuUsableBytes:V,score:q,breakdown:A,error:J,timestamp:new Date().toISOString()}},_1=({device:$,modelBytes:Z=0,kvCacheBytes:X=0}={})=>{if(!$)return{totalRequiredBytes:Z+X,fitsInGpu:!1,fitsInCpu:!1,limiting:"unknown-device"};let j=Math.max(0,Number(Z)||0)+Math.max(0,Number(X)||0),W=$.hasGpu&&j>0&&j<=$.gpuUsableBytes,H=j>0&&j<=$.cpuUsableBytes,Q="ok";if(!W&&$.hasGpu)Q="gpu-memory";if(!H)Q=W?"cpu-memory":"insufficient-memory";return{totalRequiredBytes:j,fitsInGpu:W,fitsInCpu:H,limiting:Q}},z0=async({platform:$,variant:Z=null,preferVariants:X=[],variantPreference:j=[],gpuMemoryFraction:W=D0,cpuMemoryFraction:H=b0,includeBreakdown:Q=!1,totalMemoryInBytes:J,modelBytes:N=null,kvCacheBytes:G=null,limitedKvCacheBytes:_=null,dependencies:O={},defaultVariants:z=O4}={})=>{let{getBackendDevicesInfo:V,isLibVariantAvailable:R}=O;if(typeof V!=="function"||typeof R!=="function")throw TypeError("GGML capability detection requires getBackendDevicesInfo and isLibVariantAvailable functions");let q=P3({variant:Z,preferVariants:X,variantPreference:j,defaultVariants:z}),A=[];for(let L of q)try{if(!await R(L))throw Error(`Variant ${L} not available on this platform`);let E=await V(L);A.push(q4({platform:$,totalMemoryInBytes:J,variant:L,devices:E,gpuMemoryFraction:W,cpuMemoryFraction:H,ok:!0}))}catch(B){let E=B instanceof Error?B.message:String(B);A.push(q4({platform:$,totalMemoryInBytes:J,variant:L,devices:[],gpuMemoryFraction:W,cpuMemoryFraction:H,ok:!1,error:E}))}let U=A.filter((L)=>L.ok)[0]||null,Y={ok:Boolean(U),selected:U?{...U,breakdown:Q?U.breakdown:void 0}:null,attempts:A};if(!Q&&Y.selected)delete Y.selected.breakdown;if(!Y||!N&&!G)return Y;let K=(L)=>{if(!L)return L;let B=_1({device:L,modelBytes:N||0,kvCacheBytes:G||0}),E=null;if(_!=null&&_!==G)E=_1({device:L,modelBytes:N||0,kvCacheBytes:_});return{...L,fit:B,...E&&{limitedFit:E}}};return Y.selected=K(Y.selected),Y.attempts=Array.isArray(Y.attempts)?Y.attempts.map(K):Y.attempts,Y},v0="ggml-llm";var h0=S(()=>{O4=["cuda","vulkan","snapdragon","default"]});var q1="ggml-stt",L4,O1=async({platform:$,variant:Z=null,preferVariants:X=[],variantPreference:j=[],gpuMemoryFraction:W=D0,cpuMemoryFraction:H=b0,includeBreakdown:Q=!1,totalMemoryInBytes:J,modelBytes:N=null,processingBytes:G=null,kvCacheBytes:_=null,dependencies:O={}}={})=>{let z=j&&j.length>0?j:L4;return z0({platform:$,variant:Z,preferVariants:X,variantPreference:z,gpuMemoryFraction:W,cpuMemoryFraction:H,includeBreakdown:Q,totalMemoryInBytes:J,modelBytes:N,kvCacheBytes:G??_,dependencies:O,defaultVariants:L4})};var L1=S(()=>{h0();L4=["cuda","vulkan","default"]});var D3,_0=async({platform:$,totalMemoryInBytes:Z,backend:X=v0,dependencies:j,...W}={})=>{let H=D3.get(X);if(!H)throw Error(`No capability detector registered for backend "${X}"`);return await H({...W,dependencies:j,totalMemoryInBytes:Z,platform:$})};var A4=S(()=>{h0();L1();D3=new Map([[v0,z0],[q1,O1]])});var z4,A1=($)=>{let Z=$?String($).toLowerCase():"f16";return z4[Z]||z4.f16},z1=($,Z,X,j,W,H={},{totalLayers:Q=null,swaLayers:J=0,swaContext:N=null,swaContextMultiplier:G=1,swaAdditionalTokens:_=0,swaFull:O=!1}={})=>{if(!$||!Z||!X||!j||!W)return 0;let z=Q!=null&&Q!==void 0?Number(Q):Number($),V=Math.max(0,Math.floor(z));if(!V)return 0;let R=A1(H.k),q=A1(H.v),A=Number(X)*(Number(j)*R+Number(W)*q);if(!A)return 0;let w=Math.max(0,Number(Z)||0),U=Math.min(V,Math.max(0,Math.floor(Number(J)||0))),Y=Math.max(0,V-U),K=N!=null&&Number.isFinite(Number(N))?Math.max(0,Number(N)):w,L=Math.max(1,Number(G)||1),B=Math.max(0,Number(_)||0),E=K*L+B,M=O?w:Math.min(w,E),F=Y*w+U*Math.max(0,Math.floor(M));return Math.round(A*F)},C0=({modelBytes:$=0,audioLengthSeconds:Z=30,sampleRate:X=16000,bytesPerSample:j=4}={})=>{let W=Math.max(0,Number($)||0),H=Math.max(0,Math.floor(Math.max(0,Z)*X*j)),Q=1048576,J=1073741824,N;if(W<209715200)N=125829120;else if(W<524288000)N=146800640;else if(W<2147483648)N=157286400;else N=167772160;let G;if(W<209715200)G=73400320;else if(W<524288000)G=141557760;else if(W<2147483648)G=230686720;else G=230686720;let _;if(W<104857600)_=20971520;else if(W<209715200)_=31457280;else if(W<524288000)_=89128960;else if(W<2147483648)_=225443840;else _=377487360;let O=N+G+_,z=W+O+H;return{modelBytes:W,audioBufferBytes:H,processingBufferBytes:O,totalBytes:z}};var B1=S(()=>{z4={f16:2,f32:4,q8_0:1,q6_k:0.75,q5_k:0.625,q5_k_m:0.625,q5_k_s:0.625,q5_1:0.625,q5_0:0.625,q4_k:0.5,q4_k_m:0.5,q4_k_s:0.5,q4_1:0.5,q4_0:0.5,iq4_nl:0.5}});var w1=($)=>$?String($).trim().toLowerCase():null,b3=($={},Z=null)=>{if(!$)return null;let X=w1(Z),j=X?`${X}.attention.sliding_window`:null,W=(j&&$[j]!=null?$[j]:null)??$["llama.attention.sliding_window"];if(W==null)return null;let H=Number(W);return Number.isFinite(H)?H:null},w4=($=0,Z=0,X=!1)=>{let j=Math.max(0,Math.floor(Number($)||0)),W=Math.max(0,Math.floor(Number(Z)||0));if(!j||W===1)return 0;if(W<=0)return j;let H=Math.max(0,W-1),Q=Math.floor(j/W),J=j%W,N=X?Math.max(0,J-1):Math.min(J,H);return Q*H+N},B4=({arch:$,nLayer:Z=0})=>({arch:w1($),enabled:!1,window:null,pattern:null,denseFirst:!1,type:null,kvLayers:Math.max(0,Math.floor(Number(Z)||0)),swaLayers:0}),v3,I0=({arch:$,metadata:Z={},nLayer:X=0}={})=>{let j=w1($||Z["general.architecture"]),W=Math.max(0,Math.floor(Number(X)||0)),H=b3(Z,j),Q=j?v3.get(j):null;if(!Q)return B4({arch:j,nLayer:X});let J=Q({nLayer:W,nSwa:H,metadata:Z});if(!J||!J.enabled||!J.window||J.window<=0)return B4({arch:j,nLayer:X});let N=Math.max(0,Math.floor(Number(J.pattern)||0)),G=J.kvLayers!=null&&Number.isFinite(Number(J.kvLayers))?Number(J.kvLayers):W,_=Math.max(0,Math.floor(G)),O=w4(_,N,Boolean(J.denseFirst));return{arch:j,enabled:O>0,window:J.window,pattern:N,denseFirst:Boolean(J.denseFirst),type:J.type||"standard",kvLayers:_,swaLayers:O}};var R4=S(()=>{v3=new Map([["llama4",({nSwa:$})=>{if($===0)return{enabled:!1};return{enabled:!0,window:$&&$>0?$:8192,pattern:4,type:"chunked"}}],["afmoe",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:4,type:"standard"}}],["phi3",()=>({enabled:!1})],["gemma2",({nSwa:$})=>{let Z=$&&$>0?$:4096;if(!Z)return{enabled:!1};return{enabled:!0,window:Z,pattern:2,type:"standard"}}],["gemma3",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:6,type:"standard"}}],["gemma3n",({nLayer:$,nSwa:Z})=>{if(!Z||Z<=0)return{enabled:!1};return{enabled:!0,window:Z,pattern:5,type:"standard",kvLayers:Math.min(20,$)}}],["gemma-embedding",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:6,type:"symmetric"}}],["cohere2",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:4,type:"standard"}}],["olmo2",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:4,type:"standard"}}],["exaone4",({nLayer:$,nSwa:Z})=>{let X=$>=64,j=null;if(Z&&Z>0)j=Z;else if(X)j=4096;if(!j)return{enabled:!1};return{enabled:!0,window:j,pattern:4,type:"standard"}}],["gpt-oss",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:$,pattern:2,type:"standard"}}],["smallthinker",({nSwa:$})=>{if(!$||$<=0)return{enabled:!1};return{enabled:!0,window:4096,pattern:4,denseFirst:!0,type:"standard"}}]])});var y0=($={})=>{let Z=$["general.architecture"],X=(z,V=null)=>{let R=$[z],q=Number(R);return Number.isFinite(q)?q:V},j=Z?X(`${Z}.context_length`,X("llama.context_length")):null,W=Z?X(`${Z}.block_count`,X("llama.block_count")):null,H=Z?X(`${Z}.embedding_length`,X("llama.embedding_length")):null,Q=Z?X(`${Z}.attention.head_count`,X("llama.attention.head_count")):null,J=Z?X(`${Z}.attention.head_count_kv`,X("llama.attention.head_count_kv",Q)):null,N=Z?X(`${Z}.attention.key_length`,X("llama.attention.key_length")):null,G=Z?X(`${Z}.attention.value_length`,X("llama.attention.value_length")):null,_=$["general.quantization_version"]||null,O=$["general.file_type"]||null;return{arch:Z,nCtxTrain:j,nLayer:W,nEmbd:H,nHead:Q,nHeadKv:J,nEmbdHeadK:N,nEmbdHeadV:G,quantVersion:_,fileType:O}},N0=({layerCount:$,headKvCount:Z,embdHeadKCount:X,embdHeadVCount:j,cacheTypes:W,swaConfig:H,kvUnified:Q=!1,nParallel:J=1,swaFull:N=!1})=>{let G=H?.window&&Q?Math.max(1,Number(J)||1):1;return(_)=>z1($,_,Z,X,j,W,{totalLayers:$,swaLayers:H?.swaLayers||0,swaContext:H?.window,swaFull:N,swaContextMultiplier:G})},u0=({maxCtx:$,availableMemory:Z,modelBytes:X,kvBytesForCtx:j})=>{let W=Math.max(1,Math.floor(Number($)||0));if(!j||Z<=X)return W;let H=1,Q=W,J=W;while(H<=Q){let N=Math.floor((H+Q)/2);if(X+j(N)<=Z)J=N,H=N+1;else Q=N-1}return J};var E4=S(()=>{B1()});var B0=S(()=>{A4();B1();h0();L1();R4();E4()});import{EventEmitter as h3}from"node:events";class E1{constructor($=C3){this.maxEntries=$,this.modelLoads=[],this.completions=[],this.transcriptions=[]}addModelLoad($){R1(this.modelLoads,$,this.maxEntries),o.emit("status:modelLoad",$),o.emit("status:change",{type:"modelLoad",entry:$})}addCompletion($){R1(this.completions,$,this.maxEntries),o.emit("status:completion",$),o.emit("status:change",{type:"completion",entry:$})}addTranscription($){R1(this.transcriptions,$,this.maxEntries),o.emit("status:transcription",$),o.emit("status:change",{type:"transcription",entry:$})}getModelLoadHistory(){return[...this.modelLoads].reverse()}getCompletionHistory(){return[...this.completions].reverse()}getTranscriptionHistory(){return[...this.transcriptions].reverse()}clear(){this.modelLoads=[],this.completions=[],this.transcriptions=[]}}function M1($){let Z=(X)=>$(X);return o.on("status:change",Z),()=>o.off("status:change",Z)}function F4($){M4+=1;let Z=M4,X=M1($);return{subscriberId:Z,unsubscribe:X}}function F1($){let Z=[];return{generators:Array.from($.entries()).filter(([,j])=>j.type==="ggml-llm").map(([j,W])=>{let{instance:H}=W,Q=[];if(H.contexts)Q=Array.from(H.contexts.entries()).map(([J,N])=>{let G={key:J,refCount:N.refCount,hasModel:Boolean(N.context)},_=N.context.parallel.getStatus();return G.parallelStatus=_,Z.push({generatorId:j,contextKey:J,..._}),G});return{id:j,type:W.type,refCount:W.refCount,repoId:H.info?.model?.repoId||null,quantization:H.info?.model?.quantization||null,variant:H.info?.runtime?.variant||null,nCtx:H.info?.runtime?.n_ctx||null,nParallel:H.info?.runtime?.n_parallel||null,contexts:Q}}),parallelStatuses:Z,history:{modelLoads:t.getModelLoadHistory(),completions:t.getCompletionHistory()}}}function x1($){return{generators:Array.from($.entries()).filter(([,X])=>X.type==="ggml-stt").map(([X,j])=>{let{instance:W}=j,H=W.getStatus?.()||{},Q=H.queueStatus||{processing:!1,queuedCount:0};return{id:X,type:j.type,refCount:j.refCount,repoId:W.info?.model?.repoId||null,quantization:W.info?.model?.quantization||null,variant:W.info?.runtime?.variant||null,hasContext:H.hasContext||!1,contextRefCount:H.contextRefCount||0,queueStatus:Q}}),history:{modelLoads:J0.getModelLoadHistory(),transcriptions:J0.getTranscriptionHistory()}}}function x4($){return{timestamp:new Date().toISOString(),ggmlLlm:F1($),ggmlStt:x1($)}}var C3=9999,o,R1=($,Z,X)=>{if($.push({...Z,timestamp:Z.timestamp||new Date().toISOString()}),$.length>X)$.shift()},t,J0,M4=0;var m0=S(()=>{o=new h3;o.setMaxListeners(100);t=new E1,J0=new E1});import b from"node:path";import E0 from"node:os";import{stat as M0,mkdir as I3,open as y3,unlink as e,readFile as b4,writeFile as v4,rename as h4,readdir as u3}from"node:fs/promises";import{createHash as k1}from"node:crypto";import{gguf as m3}from"@huggingface/gguf";import{loadModel as p3,getBackendDevicesInfo as C4,isLibVariantAvailable as I4}from"@fugood/llama.node";import f3 from"bytes";import*as g3 from"node:stream/web";class n4{constructor($,Z){this.config=$,this.plan=Z,this.baseDir=$.runtime.cache_dir,this.enabled=$.runtime.session_cache?.enabled!==!1,this.maxSizeBytes=Q6($.runtime.session_cache?.max_size_bytes,10737418240),this.maxEntries=$.runtime.session_cache?.max_entries||1000,this.metadata={variant:Z.info?.runtime?.variant||null,n_gpu_layers:Z.info?.runtime?.n_gpu_layers||0,n_ctx:Z.info?.runtime?.n_ctx||0,modelPath:Z.localPath,cacheTypeK:Z.info?.runtime?.cache_type_k||"f16",cacheTypeV:Z.info?.runtime?.cache_type_v||"f16"},this.cacheMap=null,this.initialized=!1}async initialize(){if(!this.enabled||this.initialized)return;try{await H0(n0(this.baseDir)),await H0(g0(this.baseDir)),await H0(l4(this.baseDir)),this.cacheMap=await r3(this.baseDir),this.initialized=!0,console.log(`[SessionCache] Initialized with ${Object.keys(this.cacheMap.entries).length} entries`)}catch($){console.warn(`[SessionCache] Failed to initialize: ${$.message}`),this.enabled=!1}}async findMatchingEntry($){if(!this.enabled||!this.cacheMap)return null;let Z=X6($,this.metadata,this.cacheMap);if(Z){let{entry:X}=Z;if(!await Y6(X.stateFilePath))return console.log(`[SessionCache] Removing stale entry: ${X.id}`),delete this.cacheMap.entries[X.id],this.cacheMap.totalSize-=X.stateFileSize||0,await P1(this.cacheMap,this.baseDir).catch(()=>{}),null;return X.lastAccessedAt=new Date().toISOString(),await P1(this.cacheMap,this.baseDir).catch(()=>{}),{entry:X}}return null}async prepareCompletionOptions($,Z){if(!this.enabled)return{options:$,cacheEntry:null,promptPrefix:null};let X=await this.findMatchingEntry(Z);if(X){let{entry:j}=X;return console.log(`[SessionCache] Found matching entry: ${j.id} (${j.fullText.length} chars, loadStateSize=${j.loadStateSize})`),{options:{...$,load_state_path:j.stateFilePath},cacheEntry:j,promptPrefix:j.fullText}}return{options:$,cacheEntry:null,promptPrefix:null}}async saveCompletionState($,Z,X,j=0){if(!this.enabled)return null;let W=a3($,this.metadata);if(this.cacheMap.entries[W])return console.log(`[SessionCache] Entry already exists for prompt: ${W}`),await e(X).catch(()=>{}),this.cacheMap.entries[W];let H=$+Z,Q=o3(W,this.baseDir);try{await H0(b.dirname(Q)),await h4(X,Q);let J=await M0(Q),N={id:W,promptText:$,completionText:Z,fullText:H,promptTokenCount:j,stateFilePath:Q,stateFileSize:J.size,metadata:{...this.metadata},createdAt:new Date().toISOString(),lastAccessedAt:new Date().toISOString()};return this.cacheMap.entries[W]=N,this.cacheMap.totalSize+=J.size,await j6(this.cacheMap,this.maxSizeBytes,this.maxEntries),await P1(this.cacheMap,this.baseDir),console.log(`[SessionCache] Saved entry: ${W} (${J.size} bytes)`),N}catch(J){return console.warn(`[SessionCache] Failed to save state: ${J.message}`),await e(X).catch(()=>{}),null}}async generateTempStatePath(){return await H0(g0(this.baseDir)),t3(this.baseDir)}async cleanup(){await W6(this.baseDir)}}async function a4($,Z,X={}){let{globalDownloadManager:j=null}=X,W=l0(Z),H=await J6(W),Q=new n4(W,H);await Q.initialize();let J={id:$,type:"ggml-llm",config:W,plan:H,info:H.info,contexts:new Map,downloads:new Map,globalDownloadManager:j,sessionCache:Q,finalized:!1},N=async()=>{if(J.finalized)return;J.finalized=!0;let Y=Array.from(J.contexts.values()),K=Y.map((B)=>{if(B.released)return Promise.resolve(!1);if(B.releaseRequested||B.releaseTimer)return Promise.resolve(!1);if(B.refCount=Math.max(0,B.refCount-1),B.refCount>0)return Promise.resolve(!1);return f0(J,B)});if(await Promise.allSettled(K),Y.length===0||Y.every((B)=>B.released))await J.sessionCache.cleanup()},G=async(Y={})=>{let{onProgress:K}=Y,L=await U6(J,K);return{modelInfo:L.modelInfo?{...L.modelInfo}:null,runtime:{...J.plan.info.runtime},download:{...J.plan.info.download}}},_=async()=>{if(J.finalized)return!1;let Y=O0(J),K=J.contexts.get(Y);if(!K)return!1;return G6(J,K,!1)},O=async(Y={})=>{let{options:K={},useCache:L=!0}=Y,B=O0(J),E=J.contexts.get(B);if(!E)throw Error(`Context "${B}" not initialized`);await E.ready;let M=K.prompt||"";if(!M&&K.messages){let x=await E.context.getFormattedChat(K.messages,K.chat_template||K.chatTemplate,{jinja:K.jinja??!0,tools:K.tools,parallel_tool_calls:K.parallel_tool_calls,tool_choice:K.tool_choice,enable_thinking:K.enable_thinking,add_generation_prompt:K.add_generation_prompt,now:K.now,chat_template_kwargs:K.chat_template_kwargs});M=x?.prompt||x||""}if(L&&J.sessionCache.enabled&&M){let{options:x}=await J.sessionCache.prepareCompletionOptions(K,M),k=await J.sessionCache.generateTempStatePath(),m=(await E.context.tokenize(M))?.tokens?.length||0,p={...x,save_state_path:k,save_state_size:m};console.log(`[SessionCache] save_state_size=${m} (prompt tokens)`);let n={repoId:J.plan.info.model?.repoId||null,quantization:J.plan.info.model?.quantization||null,variant:J.plan.info.runtime?.variant||null};return H6(E.context,p,J.sessionCache,M,k,m,J.id,n)}let F={repoId:J.plan.info.model?.repoId||null,quantization:J.plan.info.model?.quantization||null,variant:J.plan.info.runtime?.variant||null};return N6(E.context,K,J.id,F)},z=async(Y={})=>{let{text:K="",params:L={}}=Y,B=O0(J),E=J.contexts.get(B);if(!E)throw Error(`Context "${B}" not initialized`);await E.ready;let M=await E.context.tokenize(K,L);if(!M)return{tokens:[]};let F=Array.from(M.tokens??[]).map((x)=>Number(x));return{...M,tokens:F}},V=async(Y={})=>{let{tokens:K=[]}=Y,L=O0(J),B=J.contexts.get(L);if(!B)throw Error(`Context "${L}" not initialized`);await B.ready;let E=K.map((M)=>Number(M));return B.context.detokenize(E)},R=async(Y={})=>{let{messages:K=[],template:L,params:B}=Y,E=O0(J),M=J.contexts.get(E);if(!M)throw Error(`Context "${E}" not initialized`);return await M.ready,await M.context.getFormattedChat(K,L,B)},q=()=>Array.from(J.contexts.values()).some((Y)=>!Y.released&&(Y.releaseRequested||Y.releaseTimer||Y.refCount>0)),A=()=>{J.finalized=!1},w=()=>{let Y=[],K=Array.from(J.contexts.entries()).map(([L,B])=>{let E={key:L,refCount:B.refCount,hasModel:Boolean(B.context)},M=B.context.parallel.getStatus();return E.parallelStatus=M,Y.push({contextKey:L,...M}),E});return{id:J.id,type:J.type,repoId:J.plan.info.model?.repoId||null,quantization:J.plan.info.model?.quantization||null,variant:J.plan.info.runtime?.variant||null,nCtx:J.plan.info.runtime?.n_ctx||null,nParallel:J.plan.info.runtime?.n_parallel||null,contexts:K,parallelStatuses:Y}},U=(Y)=>{let K=Array.from(J.contexts.entries()).map(([L,B])=>B.context.parallel.subscribeToStatus((E)=>{Y({contextKey:L,...E})}));return{remove:()=>{K.forEach((L)=>{if(L?.remove)L.remove()})}}};return{id:$,type:"ggml-llm",info:H.info,contexts:J.contexts,initContext:G,completion:O,tokenize:z,detokenize:V,applyChatTemplate:R,releaseContext:_,finalize:N,getStatus:w,subscribeParallelStatus:U,hasPendingReleases:q,resetFinalized:A}}async function o4($,Z,X={}){let{onProgress:j,onComplete:W,onError:H}=X;try{let Q=l0($),J=await D1(Q),N=s4(Q,J),{repoId:G}=J;if(await R0(N,J.size)){if(console.log(`[Download] Model already exists: ${G} at ${N}`),typeof W==="function")W({localPath:N,repoId:G,alreadyExists:!0});return{started:!1,localPath:N,repoId:G,alreadyExists:!0}}let O=Z.getDownload(N);if(O)return console.log(`[Download] Already downloading: ${G}`),O.then(()=>{if(typeof W==="function")W({localPath:N,repoId:G,joinedExisting:!0})}).catch((V)=>{if(typeof H==="function")H(V)}),{started:!1,localPath:N,repoId:G,alreadyDownloading:!0};console.log(`[Download] Starting download: ${G}`);let z=(async()=>{try{if(J.isSplit&&J.splitCount>0){let V=/-(\d{5})-of-(\d{5})\.gguf$/,R=b.dirname(N),q=J.splitCount,A=0;for(let w=1;w<=q;w+=1){let U=String(w).padStart(5,"0"),Y=J.filename.replace(V,`-${U}-of-${String(q).padStart(5,"0")}.gguf`),K=`${Q.model.base_url.replace(/\/+$/,"")}/${J.repoId}/resolve/${J.revision}/${Y}`,L=b.join(R,Y);if(!await R0(L))await c0(K,J.headers,L,null,(E)=>{if(E>=0&&Number.isFinite(E)){let M=(A+E)/q;if(console.log(`[Download] ${G}: ${Math.round(M*100)}%`),typeof j==="function")j(M)}});A+=1}}else await c0(J.url,J.headers,N,J.size,(V)=>{if(V>=0&&Number.isFinite(V)){if(console.log(`[Download] ${G}: ${Math.round(V*100)}%`),typeof j==="function")j(V)}});if(console.log(`[Download] Completed: ${G}`),typeof W==="function")W({localPath:N,repoId:G})}catch(V){if(console.error(`[Download] Failed: ${G}`,V.message),typeof H==="function")H(V);throw V}finally{Z.deleteDownload(N)}})();return Z.setDownload(N,z),{started:!0,localPath:N,repoId:G}}catch(Q){if(console.error("[Download] Failed to start download:",Q.message),typeof H==="function")H(Q);return{started:!1,localPath:null,repoId:null,error:Q.message}}}async function K6($){let Z=l0($),X=await D1(Z),j=await i4(X.url,X.headers,Z.runtime.cache_dir),{arch:W,nCtxTrain:H,nLayer:Q,nEmbd:J,nHead:N,nHeadKv:G,nEmbdHeadK:_,nEmbdHeadV:O,quantVersion:z,fileType:V}=y0(j),R=Number.isFinite(Number(Q))?Number(Q):0,q=Number.isFinite(Number(J))?Number(J):0,A=Number.isFinite(Number(N))?Number(N):0,w=Number.isFinite(Number(G))?Number(G):A,U=A>0&&q>0?q/A:128,Y=_!=null&&Number.isFinite(Number(_))?Number(_):U,K=O!=null&&Number.isFinite(Number(O))?Number(O):U,L=I0({arch:W,metadata:j,nLayer:R}),B=L&&Number.isFinite(Number(L.kvLayers))?Number(L.kvLayers):R,E=Math.max(0,Math.floor(Number(B)||0)),F=(Z.model.n_ctx?Number(Z.model.n_ctx):null)||H||4096,x={k:Z.model.cache_type_k,v:Z.model.cache_type_v},k=X.size>0?X.size:0,D=N0({layerCount:E,headKvCount:w,embdHeadKCount:Y,embdHeadVCount:K,cacheTypes:x,swaConfig:L,kvUnified:Z.model.kv_unified,nParallel:Z.model.n_parallel,swaFull:Z.model.swa_full}),m=Z.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(Z.backend.gpu_memory_fraction))):w0.backend.gpu_memory_fraction||1,p=Z.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(Z.backend.cpu_memory_fraction))):d0,n=D(F),i=await r4(Z,{modelBytes:k,kvCacheBytes:n}),c=(i.selected.totalMemory||0)*m,T=Math.max(0,E0.totalmem()*p),f=i.selected.hasGpu?c:T,h=u0({maxCtx:F,availableMemory:f,modelBytes:k,kvBytesForCtx:D}),s=D(F),I=D(h);return{kvInfo:{nCtxTrain:H,nLayer:R,nEmbd:q,nHeadKv:w,nEmbdHeadK:Y,nEmbdHeadV:K,nHeadCount:A,nHeadKvCount:w,kvLayerCount:E,swa:L?.enabled?{window:L.window,pattern:L.pattern,denseFirst:L.denseFirst,type:L.type,layers:L.swaLayers}:null},modelBytes:k,kvCacheBytes:s,limitedKvCacheBytes:I,memoryLimitedCtx:h,quantization:{name:X.quantization||null,fileType:V,version:z}}}async function t4($=null,Z={}){let{threshold:X=1.1,includeBreakdown:j=!1,config:W,...H}=Z,Q=null,J=null,N=null,G=null,_=null,O=null;if(W)try{let{modelBytes:K,kvCacheBytes:L,limitedKvCacheBytes:B,memoryLimitedCtx:E,kvInfo:M,quantization:F}=await K6(W);Q=K,J=L,N=B,G=E,_=M,O=F}catch(K){}let z=W?.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(W.backend.gpu_memory_fraction))):void 0,V=W?.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(W.backend.cpu_memory_fraction))):void 0,R=await _0({...H,platform:process.platform,totalMemoryInBytes:E0.totalmem(),backend:"ggml-llm",includeBreakdown:j,gpuMemoryFraction:z,cpuMemoryFraction:V,dependencies:{getBackendDevicesInfo:C4,isLibVariantAvailable:I4},modelBytes:Q,kvCacheBytes:J,limitedKvCacheBytes:N}),q=R.selected,A=D4(q);q.modelBytes=Q||null,q.kvCacheBytes=J||null,q.memoryLimitedCtx=G||null,q.limitedKvCacheBytes=N||null,q.kvInfo=_||null,q.quantization=O||null;let w=null,U=null;if($){let K=D4($);U={...$,score:K};let L="buttress",B="buttress-higher-score";if(!R.ok)L="local",B="buttress-unavailable";else if(!K&&K!==0)L="buttress",B="missing-client-score";else{let{fit:E,limitedFit:M}=U,F=q?.fit,x=q?.limitedFit,k=E?.fitsInGpu||E?.fitsInCpu||M?.fitsInGpu||M?.fitsInCpu,D=F?.fitsInGpu||F?.fitsInCpu||x?.fitsInGpu||x?.fitsInCpu;if(k&&!D)L="local",B="client-fits-in-memory";else if(D&&!k)L="buttress",B="buttress-fits-in-memory";else if(K>A*X)L="local",B="client-better";else if(A>K*X)L="buttress",B="buttress-better";else L="either",B="comparable-scores"}w={buttressScore:A,clientScore:K,threshold:X,recommendation:L,reason:B}}if(!R.ok&&!w)w={buttressScore:A,clientScore:$?.score??null,threshold:X,recommendation:"local",reason:"buttress-unavailable"};let Y=null;if(W)Y={repoId:W.model?.repo_id||null,quantization:W.model?.quantization||null,nCtx:W.model?.n_ctx||null,cacheKType:W.model?.cache_type_k||"f16",cacheVType:W.model?.cache_type_v||"f16"};return{type:"ggml-llm",timestamp:new Date().toISOString(),buttress:R,client:U,comparison:w,modelConfig:Y}}var c3=()=>{if(typeof globalThis<"u"&&globalThis.ReadableStream&&globalThis.WritableStream)return{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream};return g3},d3,y4,l3,u4=($={},Z={})=>{return Object.entries(Z||{}).forEach(([X,j])=>{if(j&&typeof j==="object"&&!Array.isArray(j)){if(!$[X]||typeof $[X]!=="object")$[X]={};u4($[X],j)}else $[X]=j}),$},n3=".gguf",m4="https://huggingface.co",p4="https://huggingface.co/api",d,f4,d0=0.5,w0,S1=($,Z=[])=>{if(!$&&$!==0)return[...Z];if(Array.isArray($))return $.filter((X)=>X!=null);return[$]},p0=($)=>{if(!$)return null;let Z=String($).toLowerCase();if(["cuda","vulkan","snapdragon","default"].includes(Z))return Z;return null},l0=($={})=>{let Z=JSON.parse(JSON.stringify(w0));if(u4(Z,$),Z.backend.variant=p0(Z.backend.variant),Z.backend.variant_preference=Array.from(new Set(S1(Z.backend.variant_preference).map(p0).filter(Boolean))),Z.backend.variant_preference.length===0)Z.backend.variant_preference=["cuda","vulkan","snapdragon","default"];if(Z.runtime.prefer_variants=Array.from(new Set(S1(Z.runtime.prefer_variants).map(p0).filter(Boolean))),Z.model.preferred_quantizations=Array.from(new Set(S1(Z.model.preferred_quantizations||Z.model.quantizations).map((X)=>X?String(X).toLowerCase():null).filter(Boolean))),Z.model.quantization){let X=String(Z.model.quantization).toLowerCase();if(!Z.model.preferred_quantizations.includes(X))Z.model.preferred_quantizations.unshift(X)}return Z.model.n_parallel=Math.max(1,Number(Z.model.n_parallel)||4),Z.model.n_batch=Math.max(1,Number(Z.model.n_batch)||512),Z.model.base_url=Z.model.base_url||m4,Z.model.api_base=Z.model.api_base||p4,Z.runtime.cache_dir=Z.runtime.cache_dir?b.resolve(Z.runtime.cache_dir):d,Z.runtime.session_cache={...w0.runtime.session_cache,...Z.runtime.session_cache||{}},Z.runtime.context_release_delay_ms=Math.max(0,Number(Z.runtime.context_release_delay_ms)||w0.runtime.context_release_delay_ms),Z},S4=($)=>{let Z=$.toLowerCase();return f4.find((j)=>Z.includes(j))||null},i3=($)=>{let Z=[];if($.backend.variant)Z.push($.backend.variant);if($.runtime.prefer_variants.length>0)Z.push(...$.runtime.prefer_variants);return Z.push(...$.backend.variant_preference),Z.push("default"),Array.from(new Set(Z.map(p0).filter(Boolean)))},H0=async($)=>{await I3($,{recursive:!0})},s3=($=d)=>b.join($,".metadata-cache"),g4=($,Z,X=d)=>{let j=k1("sha256").update($).digest("hex");return b.join(s3(X),Z,`${j}.json`)},c4=async($,Z,X=d)=>{try{let j=g4($,Z,X),W=await b4(j,"utf-8");return console.log(`[Cache] Hit ${Z} cache:`,b.basename(j)),JSON.parse(W,(H,Q)=>{if(typeof Q==="string"&&Q.startsWith("__bigint__"))return BigInt(Q.slice(10));return Q})}catch(j){return null}},T1=async($,Z,X,j=d)=>{try{let W=g4($,Z,j);await H0(b.dirname(W)),await v4(W,JSON.stringify(X,(H,Q)=>{if(typeof Q==="bigint")return`__bigint__${Q.toString()}`;return Q}),"utf-8"),console.log(`[Cache] Wrote ${Z} cache:`,b.basename(W))}catch(W){console.warn(`[Cache] Failed to write ${Z} cache:`,W.message)}},n0=($=d)=>b.join($,".session-state-cache"),d4=($=d)=>b.join(n0($),"cache-map.json"),g0=($=d)=>b.join(n0($),"temp"),l4=($=d)=>b.join(n0($),"states"),P4=()=>({version:1,entries:{},totalSize:0}),r3=async($=d)=>{try{let Z=d4($),X=await b4(Z,"utf-8"),j=JSON.parse(X);if(!j.entries||typeof j.entries!=="object")return P4();return j}catch{return P4()}},P1=async($,Z=d)=>{let X=d4(Z),j=`${X}.tmp.${Date.now()}`;try{await H0(b.dirname(X)),await v4(j,JSON.stringify($,null,2),"utf-8"),await h4(j,X)}catch(W){throw await e(j).catch(()=>{}),W}},a3=($,Z)=>{let X=JSON.stringify({text:$,model:Z.modelPath,variant:Z.variant,n_gpu_layers:Z.n_gpu_layers,n_ctx:Z.n_ctx});return k1("sha256").update(X).digest("hex").slice(0,24)},o3=($,Z=d)=>b.join(l4(Z),`${$}.bin`),t3=($=d)=>{let Z=`${Date.now()}-${Math.random().toString(36).slice(2,10)}`;return b.join(g0($),`${Z}.bin`)},e3=($,Z)=>$.modelPath===Z.modelPath&&$.variant===Z.variant&&$.n_gpu_layers===Z.n_gpu_layers&&$.n_ctx>=Z.n_ctx&&$.cacheTypeK===Z.cacheTypeK&&$.cacheTypeV===Z.cacheTypeV,$6=($,Z)=>{let X=Math.min($.length,Z.length),j=0;while(j<X&&$[j]===Z[j])j+=1;return j},Z6=100,X6=($,Z,X)=>{let j=Object.values(X.entries);console.log(`[SessionCache] Finding match for promptText (${$.length} chars)`),console.log(`[SessionCache] Checking ${j.length} cache entries`);let H=j.filter((Q)=>e3(Q.metadata,Z)).reduce((Q,J)=>{let N=$6($,J.fullText);if(N>=Z6&&N>Q.prefixLen)return{entry:J,prefixLen:N};return Q},{entry:null,prefixLen:0});if(H.entry)return console.log(`[SessionCache] Prefix match found: ${H.entry.id} (${H.prefixLen}/${H.entry.fullText.length} chars)`),{entry:H.entry,prefixLength:H.prefixLen};return console.log("[SessionCache] No match found"),null},j6=async($,Z,X)=>{let j=Object.values($.entries).sort((J,N)=>new Date(J.lastAccessedAt)-new Date(N.lastAccessedAt)),W=$.totalSize,H=Object.keys($.entries).length,Q=j.filter((J)=>{let N=W>Z,G=H>X;if(!N&&!G)return!1;return W-=J.stateFileSize||0,H-=1,!0});return await Promise.all(Q.map(async(J)=>{await e(J.stateFilePath).catch(()=>{}),delete $.entries[J.id],console.log(`[SessionCache] Evicted entry: ${J.id}`)})),$.totalSize=Math.max(0,W),Q.map((J)=>J.id)},W6=async($=d)=>{let Z=g0($);try{let X=await u3(Z),j=Date.now(),W=3600000;await Promise.all(X.map(async(H)=>{let Q=b.join(Z,H),J=await M0(Q).catch(()=>null);if(J&&j-J.mtimeMs>3600000)await e(Q).catch(()=>{}),console.log(`[SessionCache] Cleaned up temp file: ${H}`)}))}catch{}},Y6=async($)=>{try{return await M0($),!0}catch{return!1}},Q6=($,Z)=>{if($==null)return Z;if(typeof $==="number")return $;if(typeof $==="string"){let X=f3.parse($);return X!=null?X:Z}return Z},T4=async($,Z={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let X=await fetch($,Z);if(!X.ok){let j=await X.text().catch(()=>"");throw Error(`Failed to fetch ${$}: ${X.status} ${X.statusText} ${j}`.trim())}return X.json()},k4=async($,Z={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let X=await fetch($,{...Z,method:"HEAD"});if(!X.ok)throw Error(`Failed to fetch headers for ${$}: ${X.status} ${X.statusText}`);return X},i4=async($,Z,X=d)=>{let j=JSON.stringify({url:$,headers:Z}),W=await c4(j,"range-metadata",X);if(W)return W;let H=!/^https?:/i.test($),{metadata:Q}=await m3($,{fetch,additionalFetchHeaders:Z,allowLocalFile:H});return await T1(j,"range-metadata",Q,X),Q},s4=($,Z)=>{if($.model.local_path)return b.resolve($.model.local_path);let X=Z.repoId.split("/"),j=b.join($.runtime.cache_dir,...X,Z.revision);return b.join(j,Z.filename)},R0=async($,Z)=>{try{let X=await M0($);if(!Z)return!0;return X.size===Z}catch(X){return!1}},c0=async($,Z,X,j,W)=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");await H0(b.dirname(X));let H=await fetch($,{headers:Z});if(!H.ok||!H.body)throw Error(`Failed to download ${$}: ${H.status} ${H.statusText}`);let Q=await y3(X,"w"),J=Number(H.headers.get("content-length"))||j||0,N=0,G=0.05;try{await H.body.pipeTo(new l3({async write(_){if(await Q.write(_),N+=_.byteLength,typeof W==="function"&&J>0){let O=Math.min(1,N/J);while(O>=G)W(G),G+=0.05}},async close(){if(await Q.close(),typeof W==="function")W(1)},async abort(_){throw await Q.close().catch(()=>{}),await e(X).catch(()=>{}),_}}))}catch(_){throw await Q.close().catch(()=>{}),await e(X).catch(()=>{}),_}if(j){let _=await M0(X);if(_.size!==j)throw await e(X).catch(()=>{}),Error(`Downloaded file size mismatch, expected ${j} got ${_.size}`)}},D1=async($)=>{let Z=$.model.repo_id||$.model.repository||$.model.model;if(!Z)throw Error("`model.repo_id` is required in Buttress backend config");let X=$.model.revision||"main",j=$.runtime.cache_dir,W=JSON.stringify({repoId:Z,revision:X,filename:$.model.filename,url:$.model.url,quantization:$.model.quantization,preferred_quantizations:$.model.preferred_quantizations}),H=await c4(W,"artifact-info",j);if(H)return H;let Q={...$.runtime.http_headers||{}};if($.runtime.huggingface_token)Q.Authorization=`Bearer ${$.runtime.huggingface_token}`;if($.model.url){let Y=await k4($.model.url,{headers:Q}),K=Number(Y.headers.get("content-length"))||null,L=$.model.filename||$.model.url.split("/").pop(),B={repoId:Z,revision:X,filename:L,url:$.model.url,size:K,headers:Q};return await T1(W,"artifact-info",B,j),B}let{filename:J}=$.model,N=$.model.quantization&&String($.model.quantization).toLowerCase(),G=await T4(`${$.model.api_base}/models/${Z}?revision=${X}&blobs=true`,{headers:Q}),O=(G?.siblings||G?.files||[]).map((Y)=>Y.rfilename||Y.path||Y.filename).filter((Y)=>typeof Y==="string"&&Y.endsWith(n3));if(O.length===0)throw Error(`No GGUF artifacts found in repo ${Z}`);let z=$.model.preferred_quantizations.length>0?$.model.preferred_quantizations:f4,V=()=>{let Y=z.find((K)=>{return O.find((B)=>B.toLowerCase().includes(K))});if(Y)return{filename:O.find((L)=>L.toLowerCase().includes(Y)),quantization:Y};return null};if(!J){let Y=V()||{filename:O[0],quantization:null},{filename:K,quantization:L}=Y;J=K,N=L||S4(J)}else if(!N)N=S4(J);let R=`${$.model.base_url.replace(/\/+$/,"")}/${Z}/resolve/${X}/${J}`,q=/-(\d{5})-of-(\d{5})\.gguf$/,A=J.match(q),w=null;if(A){let[,,Y]=A,K=await T4(`${$.model.api_base}/models/${Z}?revision=${X}&blobs=true`,{headers:Q}),L=K?.siblings||K?.files||[],B=Number(Y);w=0;for(let E=1;E<=B;E+=1){let M=String(E).padStart(5,"0"),F=J.replace(q,`-${M}-of-${Y}.gguf`),x=L.find((D)=>(D.rfilename||D.path||D.filename)===F),k=Number(x?.size);if(Number.isFinite(k)&&k>0)w+=k}}else{let Y=await k4(R,{headers:Q});w=Number(Y.headers.get("content-length"))||null}let U={repoId:Z,revision:X,filename:J,url:R,size:w,quantization:N,headers:Q,isSplit:Boolean(A),splitCount:A?Number(A[2]):0};return await T1(W,"artifact-info",U,j),U},r4=async($,{modelBytes:Z=null,kvCacheBytes:X=null}={})=>{let j=i3($),[W,...H]=j,Q=$.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.gpu_memory_fraction))):w0.backend.gpu_memory_fraction||1,J=$.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.cpu_memory_fraction))):d0,N=await _0({platform:process.platform,totalMemoryInBytes:E0.totalmem(),backend:"ggml-llm",variant:W||null,preferVariants:H,gpuMemoryFraction:Q,cpuMemoryFraction:J,dependencies:{getBackendDevicesInfo:C4,isLibVariantAvailable:I4},modelBytes:Z,kvCacheBytes:X}),G=(z)=>({...z,devices:Array.isArray(z.devices)?z.devices:[],ok:z.ok,hasGpu:Boolean(z.hasGpu),totalMemory:z.gpuTotalBytes||z.totalMemory||0,error:z.ok?null:Error(z.error||`Variant ${z.variant} not available on this platform`)});if(!N.ok||!N.selected){let z=(N.attempts||[]).map((V)=>`${V.variant}: ${V.error||"unknown error"}`).join("; ");throw Error(`Unable to initialize any backend variant (${j.join(", ")}). Errors: ${z}`)}let _=(N.attempts||[]).map(G);return{selected:G(N.selected),attempts:_}},J6=async($)=>{let Z=await D1($),X=await i4(Z.url,Z.headers,$.runtime.cache_dir),{arch:j,nCtxTrain:W,nLayer:H,nEmbd:Q,nHead:J,nHeadKv:N,nEmbdHeadK:G,nEmbdHeadV:_,quantVersion:O,fileType:z}=y0(X),V=Number.isFinite(Number(H))?Number(H):0,R=Number.isFinite(Number(Q))?Number(Q):0,q=Number.isFinite(Number(J))?Number(J):0,A=Number.isFinite(Number(N))?Number(N):q,w=q>0&&R>0?R/q:128,U=G!=null&&Number.isFinite(Number(G))?Number(G):w,Y=_!=null&&Number.isFinite(Number(_))?Number(_):w,K=I0({arch:j,metadata:X,nLayer:V}),L=K&&Number.isFinite(Number(K.kvLayers))?Number(K.kvLayers):V,B=Math.max(0,Math.floor(Number(L)||0)),E={use_mmap:$.model.use_mmap??$.runtime.use_mmap,use_mlock:$.model.use_mlock??$.runtime.use_mlock,n_threads:$.model.n_threads??$.runtime.n_threads,n_ctx:$.model.n_ctx??$.runtime.n_ctx,n_batch:$.model.n_batch??$.runtime.n_batch,n_ubatch:$.model.n_ubatch??$.runtime.n_ubatch,n_cpu_moe:$.model.n_cpu_moe??$.runtime.n_cpu_moe,n_parallel:$.model.n_parallel??$.runtime.n_parallel,cpu_mask:$.model.cpu_mask??$.runtime.cpu_mask,cpu_strict:$.model.cpu_strict??$.runtime.cpu_strict,devices:$.model.devices??$.runtime.devices,n_gpu_layers:$.model.n_gpu_layers??$.runtime.n_gpu_layers,flash_attn_type:$.model.flash_attn_type??$.runtime.flash_attn_type,cache_type_k:$.model.cache_type_k??$.runtime.cache_type_k,cache_type_v:$.model.cache_type_v??$.runtime.cache_type_v,kv_unified:$.model.kv_unified??$.runtime.kv_unified,swa_full:$.model.swa_full??$.runtime.swa_full,ctx_shift:$.model.ctx_shift??$.runtime.ctx_shift},M=E.n_ctx?Number(E.n_ctx):null,F=M||W||4096,x=[],k=[],D=!0;if(M&&W&&M>W){D=!1;let Q0=`Requested context length (${M}) exceeds model training context (${W})`;x.push(Q0),k.push(Q0),F=W}if(M&&!W)x.push("Model metadata missing training context length, using requested value");let m={k:E.cache_type_k,v:E.cache_type_v},p=Z.size>0?Z.size:0,n=N0({layerCount:B,headKvCount:A,embdHeadKCount:U,embdHeadVCount:Y,cacheTypes:m,swaConfig:K,kvUnified:E.kv_unified,nParallel:E.n_parallel,swaFull:E.swa_full}),i=n(F),v=await r4($,{modelBytes:p,kvCacheBytes:i}),c=v.selected.totalMemory||0,T=c*($.backend.gpu_memory_fraction||1),f=$.backend.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.cpu_memory_fraction))):d0,h=Math.max(0,E0.totalmem()*f),s=v.selected.hasGpu?T:h,I=u0({maxCtx:F,availableMemory:s,modelBytes:p,kvBytesForCtx:n});if(!M&&I){let Q0=W?Math.min(I,W):I,U1=Math.max(32,Q0);if(U1<F)x.push(`Context length capped to ${U1} by memory limits`);F=U1}if(F>I)F=I;let a=Math.floor(I);console.log(`[buttress] Memory-limited context length: ${a}`);let q0=n(F),T0=p+q0,G0=V?p/(V+1):p,k0=0;if(v.selected.hasGpu&&G0>0)k0=Math.min(V+1,Math.max(0,Math.floor(T/G0)));console.log(`[buttress] Auto GPU layer capacity (${v.selected.variant}): ${k0}/${V+1}`);let H1;if(E.n_gpu_layers==="auto"||E.n_gpu_layers==null)H1=k0;else H1=Math.max(0,Math.min(Number(E.n_gpu_layers)||0,V+1));let w3=(()=>{let Q0=E.flash_attn_type&&String(E.flash_attn_type).toLowerCase();if(Q0==="on"||Q0==="off")return Q0;if(Q0==="auto")return v.selected.hasGpu?"auto":"off";return v.selected.hasGpu?"auto":"off"})(),R3=$.runtime.cache_dir,V1=s4($,Z),Y4=await R0(V1,Z.size),E3={ok:D,backend:"ggml-llm",warnings:x,errors:k,model:{repoId:Z.repoId,revision:Z.revision,filename:Z.filename,quantization:Z.quantization,url:Z.url,sizeBytes:Z.size,metadata:{architecture:j,n_ctx_train:W,n_layer:V,n_embd:R,quantization_version:O,file_type:z,kv_layer_count:B,swa:K?.enabled?{window:K.window,pattern:K.pattern,dense_first:K.denseFirst,type:K.type,layers:K.swaLayers}:null}},runtime:{...E,variant:v.selected.variant,n_ctx:F,requested_ctx:M,n_gpu_layers:H1,auto_gpu_layers:k0,flash_attn_type:w3,cache_type_k:m.k,cache_type_v:m.v,estimated_max_n_ctx:a},resources:{modelBytes:p,kvCacheBytes:q0,totalEstimatedBytes:T0,gpuCapacityBytes:c,gpuUsableBytes:T,cpuUsableBytes:h,fit:v.selected.fit},devices:{selected:v.selected,attempts:v.attempts},download:{cacheDir:R3,localPath:V1,exists:Y4},timestamp:new Date().toISOString()};return{config:$,info:E3,artifact:Z,metadata:{arch:j,nCtxTrain:W,nLayer:V,nEmbd:R},devices:v,cacheTypes:m,localPath:V1,localExists:Y4}},N6=($,Z,X=null,j=null)=>{let W,H=Date.now(),Q=0;return new y4({async start(J){try{let N=await $.parallel.completion(Z,(V,R)=>{if(!R)return;if(R.token)Q+=1;J.enqueue({event:"token",data:{requestId:V,...R}})}),{requestId:G}=N;W=N.stop;let _=await N.promise;console.log("[Completion] Result:",_),J.enqueue({event:"result",data:{requestId:G,..._}}),J.close();let O=Date.now()-H,z=_.timings||{};t.addCompletion({id:`completion-${G}`,generatorId:X,requestId:G,repoId:j?.repoId||null,quantization:j?.quantization||null,variant:j?.variant||null,promptTokens:z.prompt_n??0,tokensGenerated:z.predicted_n??Q,tokensPerSecond:z.predicted_per_second??0,promptPerSecond:z.prompt_per_second??0,durationMs:O,success:!0,interrupted:_.interrupted||!1,contextFull:_.context_full||_.contextFull||!1})}catch(N){J.enqueue({event:"error",data:{message:N?.message||String(N)}}),J.error(N),t.addCompletion({id:`completion-${Date.now()}`,generatorId:X,repoId:j?.repoId||null,quantization:j?.quantization||null,variant:j?.variant||null,durationMs:Date.now()-H,tokensGenerated:Q,success:!1,error:N?.message||String(N)})}},cancel(){if(W)W()}})},H6=($,Z,X,j,W,H,Q=null,J=null)=>{let N,G="",_=!1,O=Date.now(),z=0;return new y4({async start(V){try{let R=await $.parallel.completion(Z,(Y,K)=>{if(!K)return;if(K.token)G+=K.token,z+=1;V.enqueue({event:"token",data:{requestId:Y,...K}})}),{requestId:q}=R;N=R.stop;let A=await R.promise;if(A.text)G=A.text;else if(A.content)G=A.content;_=!A.interrupted&&!A.context_full,console.log("[Completion] Result:",A),V.enqueue({event:"result",data:{requestId:q,...A}}),V.close();let w=Date.now()-O,U=A.timings||{};if(t.addCompletion({id:`completion-${q}`,generatorId:Q,requestId:q,repoId:J?.repoId||null,quantization:J?.quantization||null,variant:J?.variant||null,promptTokens:U.prompt_n??H??0,tokensGenerated:U.predicted_n??z,tokensPerSecond:U.predicted_per_second??0,promptPerSecond:U.prompt_per_second??0,durationMs:w,success:!0,interrupted:A.interrupted||!1,contextFull:A.context_full||A.contextFull||!1,usedCache:Boolean(Z.load_state_path)}),_&&X.enabled&&G)X.saveCompletionState(j,G,W,H).catch((Y)=>{console.warn("[SessionCache] Save failed:",Y.message)});else if(W)e(W).catch(()=>{})}catch(R){V.enqueue({event:"error",data:{message:R?.message||String(R)}}),V.error(R),t.addCompletion({id:`completion-${Date.now()}`,generatorId:Q,repoId:J?.repoId||null,quantization:J?.quantization||null,variant:J?.variant||null,durationMs:Date.now()-O,tokensGenerated:z,success:!1,error:R?.message||String(R)}),e(W).catch(()=>{})}},cancel(){if(N)N();e(W).catch(()=>{})}})},O0=($)=>{let Z={model:$.plan.localPath,runtime:$.plan.info.runtime};return k1("sha256").update(JSON.stringify(Z)).digest("hex").slice(0,24)},V6=async($,Z,X,j=null)=>{let{config:W,localPath:H,artifact:Q}=$;if($.localExists&&!Z.has(H)){if($.info.download.exists=!0,typeof X==="function")X(0.5);return H}if(W.model.local_path&&!W.model.allow_local_file)throw Error("Local model path provided but `model.allow_local_file` is not enabled");let J=H;if(j){let N=j.getDownload(J);if(N){console.log(`[ensureModelFile] Waiting for global download: ${Q.repoId}`);try{if(await N,await R0(H,Q.size)){if($.localExists=!0,$.info.download.exists=!0,typeof X==="function")X(0.5);return H}}catch(G){console.warn(`[ensureModelFile] Global download failed, will retry: ${G.message}`)}}}if(!Z.has(J))Z.set(J,(async()=>{if(Q.isSplit&&Q.splitCount>0){let N=/-(\d{5})-of-(\d{5})\.gguf$/,G=b.dirname(H),_=Q.splitCount,O=0;for(let z=1;z<=_;z+=1){let V=String(z).padStart(5,"0"),R=Q.filename.replace(N,`-${V}-of-${String(_).padStart(5,"0")}.gguf`),q=`${W.model.base_url.replace(/\/+$/,"")}/${Q.repoId}/resolve/${Q.revision}/${R}`,A=b.join(G,R);if(!await R0(A))await c0(q,Q.headers,A,null,(U)=>{if(U>=0&&Number.isFinite(U)){let Y=(O+U)/_,K=Math.round(Y*100);if(console.log(`Downloading model splits: ${Math.min(100,K)}%`),typeof X==="function")X(Y*0.5)}});O+=1}}else console.log("Downloading model: 0%"),await c0(Q.url,Q.headers,H,Q.size,(N)=>{if(N>=0&&Number.isFinite(N)){let G=Math.round(N*100);if(console.log(`Downloading model: ${Math.min(100,G)}%`),typeof X==="function")X(N*0.5)}});$.localExists=!0,$.info.download.exists=!0})());try{await Z.get(J)}finally{Z.delete(J)}return H},U6=async($,Z)=>{let X=O0($),j=$.contexts.get(X);if(j&&!j.released){if(j.releaseTimer)clearTimeout(j.releaseTimer),j.releaseTimer=null,console.log(`[Context] Cancelled pending release for context "${X}"`);if(j.releaseRequested=!1,j.refCount+=1,console.log(`[Context] Reusing existing context "${X}", refCount=${j.refCount}`),typeof Z==="function")Z(0);if(!j.context)await j.ready;if(typeof Z==="function")Z(1);return j}if(j)console.log(`[Context] Record exists but released=${j.released}, creating new context`);else console.log(`[Context] No existing record for "${X}", creating new context`);j={key:X,refCount:1,ready:null,released:!1},$.contexts.set(X,j),j.ready=(async()=>{let W=Date.now(),H=await V6($.plan,$.downloads,Z,$.globalDownloadManager);if(typeof Z==="function")Z(0.5);let Q={model:H,n_threads:$.plan.info.runtime.n_threads,use_mmap:$.plan.info.runtime.use_mmap,use_mlock:$.plan.info.runtime.use_mlock,cpu_mask:$.plan.info.runtime.cpu_mask,cpu_strict:$.plan.info.runtime.cpu_strict,devices:$.plan.info.runtime.devices,n_ctx:$.plan.info.runtime.n_ctx,n_gpu_layers:$.plan.info.runtime.n_gpu_layers,n_parallel:$.plan.info.runtime.n_parallel,n_batch:$.plan.info.runtime.n_batch,n_ubatch:$.plan.info.runtime.n_ubatch,n_cpu_moe:$.plan.info.runtime.n_cpu_moe,flash_attn_type:$.plan.info.runtime.flash_attn_type,ctx_shift:$.plan.info.runtime.ctx_shift,kv_unified:$.plan.info.runtime.kv_unified,swa_full:$.plan.info.runtime.swa_full,lib_variant:$.plan.info.runtime.variant};if($.plan.info.runtime.flash_attn_type!=="off")Q.cache_type_k=$.plan.info.runtime.cache_type_k,Q.cache_type_v=$.plan.info.runtime.cache_type_v;console.log("[Context] Load Options:",Q);let J;try{if(J=await p3(Q,(N)=>{if(typeof Z==="function"){if(Z(0.5+N*0.25),N%5===0)console.log("[Context] Load Model Progress:",N)}}),$.plan.info.runtime.n_parallel){if(!await J.parallel.enable({n_parallel:$.plan.info.runtime.n_parallel,n_batch:$.plan.info.runtime.n_batch}))throw Error("Failed to enable parallel decoding mode for context")}if(typeof Z==="function")Z(1);return j.context=J,j.modelInfo=J.getModelInfo(),t.addModelLoad({id:`${$.id}-${X}`,generatorId:$.id,contextKey:X,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,variant:$.plan.info.runtime?.variant||null,nCtx:$.plan.info.runtime?.n_ctx||null,nGpuLayers:$.plan.info.runtime?.n_gpu_layers||null,durationMs:Date.now()-W,success:!0}),j}catch(N){if(t.addModelLoad({id:`${$.id}-${X}`,generatorId:$.id,contextKey:X,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,variant:$.plan.info.runtime?.variant||null,durationMs:Date.now()-W,success:!1,error:N?.message||String(N)}),J)try{J.release()}catch(G){}throw N}})();try{return await j.ready,j}catch(W){throw $.contexts.delete(X),W}},f0=async($,Z,X=!1)=>{if(Z.released)return!1;if(!X&&Z.refCount>0)return!1;Z.released=!0,$.contexts.delete(Z.key);try{Z.context?.parallel?.disable?.()}catch(j){}return await Z.context?.release?.(),!0},G6=async($,Z,X=!1)=>{if(Z.releaseRequested=!0,Z.releaseTimer)clearTimeout(Z.releaseTimer),Z.releaseTimer=null;if(X)Z.refCount=0;else if(Z.refCount=Math.max(0,Z.refCount-1),Z.refCount>0)return Z.releaseRequested=!1,!1;let j=$.config.runtime.context_release_delay_ms;if(typeof j!=="number"||!Number.isFinite(j))return f0($,Z);let W=Math.max(0,Math.floor(j));if(X||W<=0)return f0($,Z);return console.log(`[Context] Scheduling release in ${W}ms for context "${Z.key}"`),Z.releaseTimer=setTimeout(async()=>{if(Z.releaseTimer=null,Z.refCount>0){console.log(`[Context] Release cancelled, refCount=${Z.refCount} for context "${Z.key}"`),Z.releaseRequested=!1;return}console.log(`[Context] Releasing context "${Z.key}" after ${W}ms delay`),await f0($,Z)},W),!0},b1=($)=>{let Z=l0($);return Z.model.repo_id||Z.model.repository||Z.model.model||null},D4=($)=>{if(!$)return 0;if(typeof $.score==="number"&&Number.isFinite($.score))return Number($.score);return K0($)};var v1=S(()=>{B0();m0();d3=c3(),{ReadableStream:y4,WritableStream:l3}=d3,d=b.join(E0.homedir(),".buttress","models"),f4=["mxfp4","q8_0","q6_k","q6","q5_k_m","q5_k_s","q5_k","q5_1","q5_0","q4_k_m","q4_k_s","q4_k","q4_1","q4_0","q3","q2"],w0={backend:{type:"ggml-llm",variant:null,variant_preference:["cuda","vulkan","snapdragon","default"],gpu_memory_fraction:0.85,cpu_memory_fraction:d0},model:{repo_id:null,revision:"main",filename:null,url:null,quantization:null,preferred_quantizations:[],n_ctx:null,n_gpu_layers:"auto",allow_local_file:!1,local_path:null,api_base:p4,base_url:m4},runtime:{cache_dir:d,prefer_variants:[],huggingface_token:process.env.HUGGINGFACE_TOKEN||null,http_headers:{},session_cache:{enabled:!0,max_size_bytes:10737418240,max_entries:1000},context_release_delay_ms:1e4}}});import X0 from"node:path";import p1 from"node:os";import{stat as X2,mkdir as _6,open as q6,unlink as h1,readFile as O6,writeFile as L6}from"node:fs/promises";import{createHash as A6}from"node:crypto";import{initWhisper as z6}from"@fugood/whisper.node";import{getBackendDevicesInfo as j2,isLibVariantAvailable as W2}from"@fugood/llama.node";import*as B6 from"node:stream/web";class _2{constructor(){this.queue=[],this.processing=!1,this.currentTaskId=null}async enqueue($,Z=null){return new Promise((X,j)=>{this.queue.push({task:$,resolve:X,reject:j,taskId:Z}),this.processNext()})}async processNext(){if(this.processing||this.queue.length===0)return;this.processing=!0;let{task:$,resolve:Z,reject:X,taskId:j}=this.queue.shift();this.currentTaskId=j;try{let W=await $();Z(W)}catch(W){X(W)}finally{this.processing=!1,this.currentTaskId=null,this.processNext()}}getStatus(){return{processing:this.processing,queuedCount:this.queue.length,currentTaskId:this.currentTaskId}}}async function q2($,Z,X={}){let{globalDownloadManager:j=null}=X,W=a0(Z),H=await T6(W),Q={id:$,type:"ggml-stt",config:W,plan:H,info:H.info,contextRecord:null,downloads:new Map,globalDownloadManager:j,queue:new _2,finalized:!1},J=async()=>{if(Q.finalized)return;Q.finalized=!0;let A=Q.contextRecord;if(!A)return;if(A.released)return;if(A.releaseRequested||A.releaseTimer)return;if(A.refCount=Math.max(0,A.refCount-1),A.refCount>0)return;await s0(Q,A)},N=async(A={})=>{let{onProgress:w}=A;try{let U=await b6(Q,w);return{modelInfo:U.modelInfo&&typeof U.modelInfo==="object"?{...U.modelInfo}:null,runtime:{...Q.plan.info.runtime},download:{...Q.plan.info.download}}}catch(U){throw console.error("[Context] Error initializing context:",U),U}},G=async()=>{if(Q.finalized)return!1;let A=Q.contextRecord;if(!A)return!1;return v6(Q,A)},_=async(A={})=>{let{audioPath:w,audioData:U,options:Y={}}=A,K=Q.contextRecord;if(!K)throw Error("Context not initialized");let L={...Y};if(Q.plan.info.runtime.max_threads&&L.maxThreads==null)L.maxThreads=Q.plan.info.runtime.max_threads;let B=`transcription-${Date.now()}-${Math.random().toString(36).slice(2,8)}`,E=Date.now();return Q.queue.enqueue(async()=>{await K.ready;try{let M;if(U){let F=D6(U),{promise:x}=K.context.transcribeData(F,L);M=await x}else{if(!w)throw Error("audioPath or audioData is required for transcription");let F=X0.resolve(w),{promise:x}=K.context.transcribe(F,L);M=await x}return J0.addTranscription({id:B,generatorId:Q.id,repoId:Q.plan.info.model?.repoId||null,quantization:Q.plan.info.model?.quantization||null,variant:Q.plan.info.runtime?.variant||null,durationMs:Date.now()-E,segmentCount:M?.segments?.length||0,textLength:M?.text?.length||0,success:!0}),M}catch(M){throw J0.addTranscription({id:B,generatorId:Q.id,repoId:Q.plan.info.model?.repoId||null,quantization:Q.plan.info.model?.quantization||null,variant:Q.plan.info.runtime?.variant||null,durationMs:Date.now()-E,success:!1,error:M?.message||String(M)}),M}},B)},O=async(A={})=>_(A),z=async(A={})=>_(A),V=()=>{let A=Q.contextRecord;if(!A)return!1;return!A.released&&(A.releaseRequested||A.releaseTimer||A.refCount>0)},R=()=>{Q.finalized=!1},q=()=>({id:Q.id,type:Q.type,repoId:Q.plan.info.model?.repoId||null,quantization:Q.plan.info.model?.quantization||null,variant:Q.plan.info.runtime?.variant||null,hasContext:Boolean(Q.contextRecord?.context),contextRefCount:Q.contextRecord?.refCount||0,queueStatus:Q.queue.getStatus()});return{id:$,type:"ggml-stt",info:H.info,queue:Q.queue,initContext:N,transcribe:O,transcribeData:z,releaseContext:G,finalize:J,getStatus:q,hasPendingReleases:V,resetFinalized:R}}async function O2($,Z,X={}){let{onProgress:j,onComplete:W,onError:H}=X;try{let Q=a0($),J=await f1(Q),N=G2(Q,J),{repoId:G}=J;if(await r0(N,J.size)){if(console.log(`[Download] STT model already exists: ${G} at ${N}`),typeof W==="function")W({localPath:N,repoId:G,alreadyExists:!0});return{started:!1,localPath:N,repoId:G,alreadyExists:!0}}let O=Z.getDownload(N);if(O)return console.log(`[Download] Already downloading STT model: ${G}`),O.then(()=>{if(typeof W==="function")W({localPath:N,repoId:G,joinedExisting:!0})}).catch((V)=>{if(typeof H==="function")H(V)}),{started:!1,localPath:N,repoId:G,alreadyDownloading:!0};console.log(`[Download] Starting STT model download: ${G}`);let z=(async()=>{try{if(await K2(J.url,J.headers,N,J.size,(V)=>{if(V>=0&&Number.isFinite(V)){if(console.log(`[Download] ${G}: ${Math.round(V*100)}%`),typeof j==="function")j(V)}}),console.log(`[Download] Completed STT model: ${G}`),typeof W==="function")W({localPath:N,repoId:G})}catch(V){if(console.error(`[Download] Failed STT model: ${G}`,V.message),typeof H==="function")H(V);throw V}finally{Z.deleteDownload(N)}})();return Z.setDownload(N,z),{started:!0,localPath:N,repoId:G}}catch(Q){if(console.error("[Download] Failed to start STT download:",Q.message),typeof H==="function")H(Q);return{started:!1,localPath:null,repoId:null,error:Q.message}}}async function L2($=null,Z={}){let{threshold:X=1.1,includeBreakdown:j=!1,config:W,...H}=Z,Q=null,J=null,N=null;if(W)try{let w=a0(W),U=await f1(w);Q=U.size??null,{processingBufferBytes:J}=C0({modelBytes:Q}),N=U.quantization||null}catch(w){}let G=W?.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(W.backend.gpu_memory_fraction))):void 0,_=W?.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number(W.backend.cpu_memory_fraction))):void 0,O=await _0({...H,platform:process.platform,totalMemoryInBytes:p1.totalmem(),backend:"ggml-stt",includeBreakdown:j,gpuMemoryFraction:G,cpuMemoryFraction:_,dependencies:{getBackendDevicesInfo:j2,isLibVariantAvailable:W2},modelBytes:Q,kvCacheBytes:J}),z=O.selected,V=Z2(z);if(z)z.modelBytes=Q||null,z.processingBytes=J||null,z.quantization=N||null;let R=null,q=null;if($){let w=Z2($);q={...$,score:w};let U="buttress",Y="buttress-higher-score";if(!O.ok)U="local",Y="buttress-unavailable";else if(!w&&w!==0)U="buttress",Y="missing-client-score";else if($.fit&&z?.fit){let K=$.fit.fitsInGpu||$.fit.fitsInCpu,L=z.fit.fitsInGpu||z.fit.fitsInCpu;if(K&&!L)U="local",Y="client-fits-in-memory";else if(L&&!K)U="buttress",Y="buttress-fits-in-memory";else if(w>V*X)U="local",Y="client-better";else if(V>w*X)U="buttress",Y="buttress-better";else U="either",Y="comparable-scores"}else if(w>V*X)U="local",Y="client-better";else if(V>w*X)U="buttress",Y="buttress-better";else U="either",Y="comparable-scores";R={buttressScore:V,clientScore:w,threshold:X,recommendation:U,reason:Y}}if(!O.ok&&!R)R={buttressScore:V,clientScore:$?.score??null,threshold:X,recommendation:"local",reason:"buttress-unavailable"};let A=null;if(W)A={repoId:W.model?.repo_id||null,quantization:W.model?.quantization||null,filename:W.model?.filename||null};return{type:"ggml-stt",timestamp:new Date().toISOString(),buttress:O,client:q,comparison:R,modelConfig:A}}var w6=()=>{if(typeof globalThis<"u"&&globalThis.ReadableStream&&globalThis.WritableStream)return{ReadableStream:globalThis.ReadableStream,WritableStream:globalThis.WritableStream};return B6},R6,Y2=($={},Z={})=>{return Object.entries(Z||{}).forEach(([X,j])=>{if(j&&typeof j==="object"&&!Array.isArray(j)){if(!$[X]||typeof $[X]!=="object")$[X]={};Y2($[X],j)}else $[X]=j}),$},E6=".bin",Q2="https://huggingface.co",J2="https://huggingface.co/api",L0,y1,u1,N2="fp16",H2=0.5,m1,C1=($,Z=[])=>{if(!$&&$!==0)return[...Z];if(Array.isArray($))return $.filter((X)=>X!=null);return[$]},i0=($)=>{if(!$)return null;let Z=String($).toLowerCase();if(["cuda","vulkan","default"].includes(Z))return Z;return null},a0=($={})=>{let Z=JSON.parse(JSON.stringify(m1));if(Y2(Z,$),Z.backend.variant=i0(Z.backend.variant),Z.backend.variant_preference=Array.from(new Set(C1(Z.backend.variant_preference||y1).map(i0).filter(Boolean))),Z.backend.variant_preference.length===0)Z.backend.variant_preference=[...y1];if(Z.runtime.prefer_variants=Array.from(new Set(C1(Z.runtime.prefer_variants).map(i0).filter(Boolean))),Z.model.preferred_quantizations=Array.from(new Set(C1(Z.model.preferred_quantizations||Z.model.quantizations).map((X)=>X?String(X).toLowerCase():null).filter(Boolean))),Z.model.quantization){let X=String(Z.model.quantization).toLowerCase();if(!Z.model.preferred_quantizations.includes(X))Z.model.preferred_quantizations.unshift(X)}return Z.model.base_url=Z.model.base_url||Q2,Z.model.api_base=Z.model.api_base||J2,Z.runtime.cache_dir=Z.runtime.cache_dir?X0.resolve(Z.runtime.cache_dir):L0,Z.runtime.context_release_delay_ms=Math.max(0,Number(Z.runtime.context_release_delay_ms)||m1.runtime.context_release_delay_ms),Z},I1=($)=>{let Z=$.toLowerCase();return u1.find((j)=>Z.includes(j))||null},M6=($)=>{let Z=[];if($.backend.variant)Z.push($.backend.variant);if($.runtime.prefer_variants.length>0)Z.push(...$.runtime.prefer_variants);return Z.push(...$.backend.variant_preference),Z.push("default"),Array.from(new Set(Z.map(i0).filter(Boolean)))},V2=async($)=>{await _6($,{recursive:!0})},F6=($=L0)=>X0.join($,".metadata-cache"),U2=($,Z,X=L0)=>{let j=A6("sha256").update($).digest("hex");return X0.join(F6(X),Z,`${j}.json`)},x6=async($,Z,X=L0)=>{try{let j=U2($,Z,X),W=await O6(j,"utf-8");return JSON.parse(W)}catch(j){return null}},e4=async($,Z,X,j=L0)=>{try{let W=U2($,Z,j);await V2(X0.dirname(W)),await L6(W,JSON.stringify(X),"utf-8")}catch(W){}},S6=async($,Z={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let X=await fetch($,Z);if(!X.ok){let j=await X.text().catch(()=>"");throw Error(`Failed to fetch ${$}: ${X.status} ${X.statusText} ${j}`.trim())}return X.json()},$2=async($,Z={})=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");let X=await fetch($,{...Z,method:"HEAD"});if(!X.ok)throw Error(`Failed to fetch headers for ${$}: ${X.status} ${X.statusText}`);return X},G2=($,Z)=>{if($.model.local_path)return X0.resolve($.model.local_path);let X=Z.repoId.split("/"),j=X0.join($.runtime.cache_dir,...X,Z.revision);return X0.join(j,Z.filename)},r0=async($,Z)=>{try{let X=await X2($);if(!Z)return!0;return X.size===Z}catch(X){return!1}},K2=async($,Z,X,j,W)=>{if(typeof fetch!=="function")throw Error("Global fetch is not available in this runtime");await V2(X0.dirname(X));let H=await fetch($,{headers:Z});if(!H.ok||!H.body)throw Error(`Failed to download ${$}: ${H.status} ${H.statusText}`);let Q=await q6(X,"w"),J=Number(H.headers.get("content-length"))||j||0,N=0,G=0.05;try{await H.body.pipeTo(new R6({async write(_){if(await Q.write(_),N+=_.byteLength,typeof W==="function"&&J>0){let O=Math.min(1,N/J);while(O>=G)W(G),G+=0.05}},async close(){if(await Q.close(),typeof W==="function")W(1)},async abort(_){throw await Q.close().catch(()=>{}),await h1(X).catch(()=>{}),_}}))}catch(_){throw await Q.close().catch(()=>{}),await h1(X).catch(()=>{}),_}if(j){let _=await X2(X);if(_.size!==j)throw await h1(X).catch(()=>{}),Error(`Downloaded file size mismatch, expected ${j} got ${_.size}`)}},f1=async($)=>{let Z=$.model.repo_id||$.model.repository||$.model.model;if(!Z)throw Error("`model.repo_id` is required in Buttress backend config");let X=$.model.revision||"main",j=$.runtime.cache_dir,W=JSON.stringify({repoId:Z,revision:X,filename:$.model.filename,url:$.model.url,quantization:$.model.quantization,preferred_quantizations:$.model.preferred_quantizations}),H=await x6(W,"artifact-info",j);if(H)return H;let Q={...$.runtime.http_headers||{}};if($.runtime.huggingface_token)Q.Authorization=`Bearer ${$.runtime.huggingface_token}`;if($.model.url){let U=await $2($.model.url,{headers:Q}),Y=Number(U.headers.get("content-length"))||null,K=$.model.filename||$.model.url.split("/").pop(),L={repoId:Z,revision:X,filename:K,url:$.model.url,size:Y,quantization:I1(K||""),headers:Q};return await e4(W,"artifact-info",L,j),L}let{filename:J}=$.model,N=$.model.quantization&&String($.model.quantization).toLowerCase(),G=await S6(`${$.model.api_base}/models/${Z}?revision=${X}&blobs=true`,{headers:Q}),O=(G?.siblings||G?.files||[]).map((U)=>U.rfilename||U.path||U.filename).filter((U)=>typeof U==="string"&&U.endsWith(E6));if(O.length===0)throw Error(`No model artifacts found in repo ${Z}`);let z=$.model.preferred_quantizations.length>0?$.model.preferred_quantizations:u1,V=()=>{for(let U of z)if(U===N2){let Y=O.find((K)=>{let L=K.toLowerCase();return!u1.some((B)=>L.includes(B))});if(Y)return{filename:Y,quantization:null}}else{let Y=O.find((K)=>K.toLowerCase().includes(U));if(Y)return{filename:Y,quantization:U}}return null};if(!J){let U=V()||{filename:O[0],quantization:null},{filename:Y,quantization:K}=U;J=Y,N=K||I1(J)}else if(!N)N=I1(J);let R=`${$.model.base_url.replace(/\/+$/,"")}/${Z}/resolve/${X}/${J}`,q=await $2(R,{headers:Q}),A=Number(q.headers.get("content-length"))||null,w={repoId:Z,revision:X,filename:J,url:R,size:A,quantization:N,headers:Q,isSplit:!1,splitCount:0};return await e4(W,"artifact-info",w,j),w},P6=async($,{modelBytes:Z=null,processingBytes:X=null}={})=>{let j=M6($),[W,...H]=j,Q=$.backend?.gpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.gpu_memory_fraction))):m1.backend.gpu_memory_fraction||1,J=$.backend?.cpu_memory_fraction!=null?Math.min(1,Math.max(0,Number($.backend.cpu_memory_fraction))):H2,N=await _0({platform:process.platform,totalMemoryInBytes:p1.totalmem(),backend:"ggml-stt",variant:W||null,preferVariants:H,variantPreference:$.backend.variant_preference,gpuMemoryFraction:Q,cpuMemoryFraction:J,dependencies:{getBackendDevicesInfo:j2,isLibVariantAvailable:W2},modelBytes:Z,kvCacheBytes:X}),G=(z)=>({...z,devices:Array.isArray(z.devices)?z.devices:[],ok:z.ok,hasGpu:Boolean(z.hasGpu),totalMemory:z.gpuTotalBytes||z.totalMemory||0,error:z.ok?null:Error(z.error||`Variant ${z.variant} not available on this platform`)});if(!N.ok||!N.selected){let z=(N.attempts||[]).map((V)=>`${V.variant}: ${V.error||"unknown error"}`).join("; ");throw Error(`Unable to initialize any backend variant (${j.join(", ")}). Errors: ${z}`)}let _=(N.attempts||[]).map(G);return{selected:G(N.selected),attempts:_}},T6=async($)=>{let Z=await f1($),X=C0({modelBytes:Z.size>0?Z.size:0}),j=await P6($,{modelBytes:X.modelBytes,processingBytes:X.processingBufferBytes}),W=j.selected.hasGpu&&(j.selected.fit?.fitsInGpu!==void 0?j.selected.fit.fitsInGpu:!0);if($.model.use_gpu===!1)W=!1;let H=$.model.use_flash_attn&&String($.model.use_flash_attn).toLowerCase(),Q;if(H==="on"||H==="true")Q=!0;else if(H==="off"||H==="false")Q=!1;else Q=W;let J=$.runtime.cache_dir,N=G2($,Z),G=await r0(N,Z.size),_={ok:!0,backend:"ggml-stt",model:{repoId:Z.repoId,revision:Z.revision,filename:Z.filename,quantization:Z.quantization,url:Z.url,sizeBytes:Z.size},runtime:{variant:j.selected.variant,use_gpu:W,use_flash_attn:Q,max_threads:$.runtime.max_threads?Number($.runtime.max_threads):null},resources:{...X,gpuCapacityBytes:j.selected.gpuTotalBytes,gpuUsableBytes:j.selected.gpuUsableBytes,cpuUsableBytes:j.selected.cpuUsableBytes,fit:j.selected.fit},devices:{selected:j.selected,attempts:j.attempts},download:{cacheDir:J,localPath:N,exists:G},timestamp:new Date().toISOString()};return{config:$,info:_,artifact:Z,memory:X,devices:j,localPath:N,localExists:G}},k6=async($,Z,X,j=null)=>{let{localPath:W,artifact:H,config:Q}=$;if($.localExists){if(typeof X==="function")X(1);return W}if(j){let G=j.getDownload(W);if(G){console.log(`[ensureModelFile] Waiting for global STT download: ${H.repoId}`);try{if(await G,await r0(W,H.size)){if($.localExists=!0,$.info.download.exists=!0,typeof X==="function")X(1);return W}}catch(_){console.warn(`[ensureModelFile] Global STT download failed, will retry: ${_.message}`)}}}let J=Z.get(W);if(J){if(await J,typeof X==="function")X(1);return W}let N=(async()=>{if(Q.model.allow_local_file){if(!await r0(W,H.size))throw Error(`Local model file not found: ${W}`);return W}return await K2(H.url,H.headers,W,H.size,X),W})();Z.set(W,N);try{return await N,W}finally{Z.delete(W)}},D6=($)=>{if(!$)return null;if($ instanceof ArrayBuffer)return $;if(ArrayBuffer.isView($))return $.buffer;if(typeof $==="string"){let Z=$.startsWith("data:")?$.split(",")[1]||"":$,X=Buffer.from(Z,"base64");return X.buffer.slice(X.byteOffset,X.byteOffset+X.byteLength)}throw Error("Unsupported audioData format, expected base64 string or ArrayBuffer")},b6=async($,Z)=>{if($.contextRecord&&!$.contextRecord.released){if($.contextRecord.releaseTimer)clearTimeout($.contextRecord.releaseTimer),$.contextRecord.releaseTimer=null,console.log("[Context] Cancelled pending STT release");if($.contextRecord.releaseRequested=!1,$.contextRecord.refCount+=1,console.log(`[Context] Reusing existing STT context, refCount=${$.contextRecord.refCount}`),typeof Z==="function")Z(0);if(!$.contextRecord.context)await $.contextRecord.ready;if(typeof Z==="function")Z(1);return $.contextRecord}if($.contextRecord)console.log(`[Context] STT record exists but released=${$.contextRecord.released}, creating new context`);else console.log("[Context] No existing STT record, creating new context");let X={refCount:1,ready:null,released:!1};$.contextRecord=X,X.ready=(async()=>{let j=Date.now();try{if(typeof Z==="function")Z(0);let W=await k6($.plan,$.downloads,Z,$.globalDownloadManager);if(typeof Z==="function")Z(0.5);let H=await z6({filePath:W,useFlashAttn:$.plan.info.runtime.flash_attn_type==="on",useGpu:$.plan.info.runtime.n_gpu_layers>0,nThreads:$.plan.info.runtime.n_threads},$.plan.info.runtime.variant);if(typeof Z==="function")Z(1);X.context=H;try{X.modelInfo=H.getModelInfo()}catch(Q){X.modelInfo=null}return J0.addModelLoad({id:$.id,generatorId:$.id,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,variant:$.plan.info.runtime?.variant||null,useGpu:$.plan.info.runtime?.use_gpu||!1,durationMs:Date.now()-j,success:!0}),X}catch(W){throw J0.addModelLoad({id:$.id,generatorId:$.id,repoId:$.plan.info.model?.repoId||null,quantization:$.plan.info.model?.quantization||null,variant:$.plan.info.runtime?.variant||null,durationMs:Date.now()-j,success:!1,error:W?.message||String(W)}),W}})();try{if(await X.ready,typeof Z==="function")Z(1);return X}catch(j){throw $.contextRecord=null,j}},s0=async($,Z,X=!1)=>{if(Z.released)return!1;if(!X&&Z.refCount>0)return!1;return Z.released=!0,$.contextRecord=null,await Z.context?.release?.(),!0},v6=async($,Z,X=!1)=>{if(Z.releaseRequested=!0,Z.releaseTimer)clearTimeout(Z.releaseTimer),Z.releaseTimer=null;if(X)Z.refCount=0;else if(Z.refCount=Math.max(0,Z.refCount-1),Z.refCount>0)return Z.releaseRequested=!1,!1;let j=$.config.runtime.context_release_delay_ms;if(typeof j!=="number"||!Number.isFinite(j))return s0($,Z);let W=Math.max(0,Math.floor(j));if(X||W<=0)return s0($,Z);return console.log(`[Context] Scheduling STT release in ${W}ms`),Z.releaseTimer=setTimeout(async()=>{if(Z.releaseTimer=null,Z.refCount>0){console.log(`[Context] STT release cancelled, refCount=${Z.refCount}`),Z.releaseRequested=!1;return}console.log(`[Context] Releasing STT context after ${W}ms delay`),await s0($,Z)},W),!0},h6,C6=($)=>{if(!$)return null;let Z=$.toLowerCase();return h6.find((X)=>Z.includes(X))||null},g1=($)=>{let Z=a0($),X=Z.model.repo_id||Z.model.repository||Z.model.model||null;if(!X)return null;let j=C6(Z.model.filename);if(j)return`${X}:${j}`;return X},Z2=($)=>{if(!$)return 0;if(typeof $.score==="number"&&Number.isFinite($.score))return Number($.score);return K0($)};var c1=S(()=>{B0();m0();({WritableStream:R6}=w6()),L0=X0.join(p1.homedir(),".buttress","models"),y1=["cuda","vulkan","default"],u1=["q8_0","q5_1","q5_0","q4_1","q4_0"],m1={backend:{type:"ggml-stt",variant:null,variant_preference:y1,gpu_memory_fraction:0.85,cpu_memory_fraction:H2},model:{repo_id:"BricksDisplay/whisper-ggml",revision:"main",filename:null,url:null,quantization:null,preferred_quantizations:["q8_0",N2,"q5_1"],allow_local_file:!1,local_path:null,api_base:J2,base_url:Q2,use_gpu:!0,use_flash_attn:"auto"},runtime:{cache_dir:L0,prefer_variants:[],huggingface_token:process.env.HUGGINGFACE_TOKEN||null,http_headers:{},max_threads:null,context_release_delay_ms:1e4}};h6=["large-v3-turbo","distil-large-v3","large-v3","large-v2","large-v1","large","distil-medium","medium.en","medium","small.en-tdrz","distil-small.en","small.en","small","base.en","base","tiny.en","tiny"]});async function r($,Z=null,X={}){if($==="ggml-llm")return t4(Z,X);if($==="ggml-stt")return L2(Z,X);throw Error(`Unknown backend type: ${$}`)}var d1=S(()=>{v1();c1()});var C;var A2=S(()=>{C={name:"@fugood/buttress-backend-core",private:!0,type:"module",version:"2.23.0-beta.36",main:"src/index.js",types:"lib/types/index.d.ts",scripts:{build:"tsc --noResolve --noCheck --declaration --emitDeclarationOnly --allowJs --outDir lib/types src/index.js"},dependencies:{"@fugood/buttress-hardware-guardrails":"^2.23.0-beta.35","@fugood/llama.node":"^1.4.11","@fugood/whisper.node":"^1.0.11","@huggingface/gguf":"^0.3.2","@iarna/toml":"^3.0.0",bytes:"^3.1.0"}}});import y from"node:os";import z2 from"node:fs";import B2 from"node:path";import{execSync as o0}from"node:child_process";import w2 from"@iarna/toml";async function E2({modelIds:$=[],defaultConfig:Z=null}={}){let X=[];if(console.log(`${C.name} v${C.version}`),console.log(`Generating model capabilities comparison...
+`),X.push(`${C.name} v${C.version}`),X.push(`## Model Capabilities Comparison
+`),!$||$.length===0)console.error("Error: No model IDs provided"),process.exit(1);try{let j=(U={},Y={})=>{let K=Array.isArray(U)?[...U]:{...U};return Object.entries(Y||{}).forEach(([L,B])=>{if(B&&typeof B==="object"&&!Array.isArray(B))K[L]=j(K[L]||{},B);else K[L]=B}),K},W=Z||{},{server:H,generators:Q=[],...J}=W,N=(U)=>j(JSON.parse(JSON.stringify(J)),U||{}),G=(U)=>{if(Array.isArray(Q)&&Q.length>0){let Y=Q.filter((K)=>K?.type==="ggml-llm");if(Y.length>0&&U){let K=Y.find((L)=>L.model?.repo_id===U);if(K)return N(K)}}return Object.keys(J).length>0?N({}):null},_=[];for(let U=0;U<$.length;U+=1){let Y=$[U];console.log(`[${U+1}/${$.length}] Analyzing ${Y}...`);let K=G(Y);K={...K||{},model:{...J.runtime,...K?.model||{},repo_id:Y}};let L=await r("ggml-llm",null,{config:K,includeBreakdown:!0});_.push({modelId:Y,capabilities:L,modelInfo:L.buttress?.selected||null,modelConfig:L.modelConfig||null})}let O=(U)=>U?(U/1024/1024/1024).toFixed(2):"N/A",z=(U)=>U?"✅":"\uD83D\uDEAB";X.push("| Model ID | Quantization | Size (GB) | Context Size | KV Cache Size (GB) | Total Required Memory (GB) | Fits GPU (Full) | Fits CPU (Full) |"),X.push("|----------|--------------|-----------|--------------|--------------------|-----------------------------|-----------------|-----------------|"),_.forEach(({modelId:U,modelInfo:Y,modelConfig:K})=>{let L=Y?.quantization?.name?.toUpperCase()||"N/A",B=O(Y?.modelBytes),E=K?.nCtx||Y?.kvInfo?.nCtxTrain||"N/A",M=N0(Y),F=Number(E),x=Y?.kvCacheBytes||(M&&Number.isFinite(F)&&F>0?M(F):M&&M(Y?.kvInfo?.nCtxTrain||0))||null,k=O(x),D=O(Y?.modelBytes&&x?Y.modelBytes+x:Y?.fit?.totalRequiredBytes),m=z(Y?.fit?.fitsInGpu),p=z(Y?.fit?.fitsInCpu);X.push(`| ${U} | ${L} | ${B} | ${E} | ${k} | ${D} | ${m} | ${p} |`);let n=Y?.memoryLimitedCtx!=null||Y?.limitedFit!=null,i=!Y?.fit?.fitsInGpu||!Y?.fit?.fitsInCpu;if(n&&i){let v=Y?.memoryLimitedCtx||E,c=Number(v),T=Y?.limitedKvCacheBytes||M&&Number.isFinite(c)&&c>0&&M(c)||null,f=O(T),h=O(Y?.modelBytes&&T?Y.modelBytes+T:Y?.limitedFit?.totalRequiredBytes),s=z(Y?.limitedFit?.fitsInGpu),I=z(Y?.limitedFit?.fitsInCpu);if(v!==E||f!==k||h!==D)X.push(`| ↳ Limited | - | ${B} | ${v} | ${f} | ${h} | ${s} | ${I} |`)}}),X.push(`
+---`),X.push(`
+### System Information`);let V=null;if(process.platform!=="win32")try{V=o0("uname -a",{encoding:"utf8"}).trim()}catch{}if(V)X.push(`- **System:** ${V}`);else X.push(`- **Hostname:** ${y.hostname()}`),X.push(`- **OS:** ${y.type()} ${y.release()}`);if(X.push(`- **Platform:** ${process.platform}`),X.push(`- **CPU Cores:** ${y.cpus().length}`),X.push(`- **Total System Memory:** ${(y.totalmem()/1024/1024/1024).toFixed(2)} GB`),_.length>0){let Y=_[0].capabilities.buttress?.selected;if(Y){let K=Y.cpuTotalBytes>0?(Y.cpuUsableBytes/Y.cpuTotalBytes*100).toFixed(0):0;if(X.push(`- **Usable CPU Memory:** ${(Y.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${K}% of ${(Y.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),Y.hasGpu){let L=Y.devices.filter((B)=>B.type==="gpu");if(L.length>0){let B=L[0];X.push(`- **GPU Backend:** ${B.backend}`),X.push(`- **GPU Name:** ${B.deviceName}`),X.push(`- **GPU Total Memory:** ${(B.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let E=Y.gpuTotalBytes>0?(Y.gpuUsableBytes/Y.gpuTotalBytes*100).toFixed(0):0;X.push(`- **GPU Usable Memory:** ${(Y.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${E}% of ${(Y.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else X.push("- **GPU:** Not available")}}X.push(`
+### Command Used`);let R=process.argv.slice(2).join(" ");if(X.push(`\`\`\`bash
+${process.argv[0]} ${process.argv[1]} ${R}
+\`\`\``),X.push(`
+### Package Information`),X.push(`- **Name:** ${C.name}`),X.push(`- **Version:** ${C.version}`),C.description)X.push(`- **Description:** ${C.description}`);if(Z&&Object.keys(Z).length>0){X.push(`
+### Configuration`),X.push("<details>"),X.push("<summary>Click to expand TOML configuration</summary>"),X.push("\n```toml");try{let U=w2.stringify(Z);X.push(U)}catch(U){X.push("# Error serializing config"),X.push(JSON.stringify(Z,null,2))}X.push("```"),X.push("</details>")}let A=`ggml-llm-model-capabilities-${new Date().toISOString().replace(/[.:]/g,"-").split("T")[0]}.md`,w=B2.join(process.cwd(),A);z2.writeFileSync(w,X.join(`
+`),"utf8"),console.log(`
+Model capabilities table saved to: ${w}`),process.exit(0)}catch(j){console.error("Failed to generate model table:",j.message),process.exit(1)}}async function M2({modelId:$=null,defaultConfig:Z=null}={}){if(console.log(`${C.name} v${C.version}`),console.log("Testing capabilities for backend: ggml-llm"),$)console.log(`Model: ${$}`);console.log("--------------------------------");try{let X=Z||{},{server:j,generators:W=[],...H}=X,Q=(V={},R={})=>{let q=Array.isArray(V)?[...V]:{...V};return Object.entries(R||{}).forEach(([A,w])=>{if(w&&typeof w==="object"&&!Array.isArray(w))q[A]=Q(q[A]||{},w);else q[A]=w}),q},J=(V)=>Q(JSON.parse(JSON.stringify(H)),V||{}),G=((V)=>{if(Array.isArray(W)&&W.length>0){let R=W.filter((q)=>q?.type==="ggml-llm");if(R.length>0){if(V){let q=R.find((A)=>A.model?.repo_id===V);if(q)return J(q)}}}if(Object.keys(H).length>0)return J({});return null})($);if($)G={...G||{},model:{...G?.model||{},repo_id:$}};let _=await r("ggml-llm",null,{config:G,includeBreakdown:!0}),O=_.buttress?.selected||null,z=_.modelConfig||null;if($||z?.repoId){console.log(`
+=== Model Information ===`);let V=$||z?.repoId;if(console.log(`Repository ID: ${V}`),z?.quantization)console.log(`Quantization: ${z.quantization}`);if(z?.nCtx)console.log(`Context Length: ${z.nCtx}`);if(O?.quantization?.name)console.log(`Model Quantization: ${O.quantization.name.toUpperCase()}`);let R=z?.cache_type_k||"f16",q=z?.cache_type_v||"f16";if(console.log(`KV Cache Type: K=${R}, V=${q}`),O?.modelBytes&&O?.kvCacheBytes){if(console.log(`Model Size: ${(O.modelBytes/1024/1024/1024).toFixed(2)} GB`),O.kvInfo)console.log(`KV Cache Size: ${(O.kvCacheBytes/1024/1024/1024).toFixed(2)} GB (KV info: ${JSON.stringify(O.kvInfo)})`);else console.log(`KV Cache Size: ${(O.kvCacheBytes/1024/1024/1024).toFixed(2)} GB`);if(console.log(`Total Required Memory: ${((O.modelBytes+O.kvCacheBytes)/1024/1024/1024).toFixed(2)} GB`),O.memoryLimitedCtx!=null){let A=O.memoryLimitedCtx,w=O.kvInfo?.nCtxTrain;if(w)console.log(`
+Memory-Limited Context: ${A} (Train: ${w})`);else console.log(`
+Memory-Limited Context: ${A}`);if(O.limitedKvCacheBytes!=null)console.log(`Limited KV Cache Size: ${(O.limitedKvCacheBytes/1024/1024/1024).toFixed(2)} GB`)}}else if(_.buttress?.selected?.fit){let{totalRequiredBytes:A}=_.buttress.selected.fit;console.log(`Total Required Memory: ${(A/1024/1024/1024).toFixed(2)} GB`)}}if(_.buttress?.selected){let{selected:V}=_.buttress;console.log(`
+=== Hardware Information ===`);let R=null;if(process.platform!=="win32")try{R=o0("uname -a",{encoding:"utf8"}).trim()}catch{}if(R)console.log(`System: ${R}`);else console.log(`Hostname: ${y.hostname()}`),console.log(`OS: ${y.type()} ${y.release()}`);console.log(`Platform: ${V.platform}`),console.log(`CPU Cores: ${y.cpus().length}`),console.log(`Total System Memory: ${(y.totalmem()/1024/1024/1024).toFixed(2)} GB`);let q=V.cpuTotalBytes>0?(V.cpuUsableBytes/V.cpuTotalBytes*100).toFixed(0):0;if(console.log(`Usable CPU Memory: ${(V.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${q}% of ${(V.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),V.hasGpu)console.log(`
+--- GPU Details ---`),V.devices.filter((w)=>w.type==="gpu").forEach((w)=>{console.log(`GPU Backend: ${w.backend}`),console.log(`GPU Name: ${w.deviceName}`),console.log(`GPU Total Memory: ${(w.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let U=V.gpuTotalBytes>0?(V.gpuUsableBytes/V.gpuTotalBytes*100).toFixed(0):0;if(console.log(`GPU Usable Memory: ${(V.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${U}% of ${(V.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),w.metadata){if(w.metadata.hasBFloat16)console.log("Supports BFloat16: Yes");if(w.metadata.hasUnifiedMemory)console.log("Unified Memory: Yes")}});else console.log("GPU: Not available");if(console.log(`
+Backend Variant: ${V.variant}`),console.log(`Performance Score: ${V.score}`),V.fit){if(console.log(`
+--- Model Fit Analysis ---`),console.log(`Fits in GPU: ${V.fit.fitsInGpu?"Yes":"No"}`),console.log(`Fits in CPU: ${V.fit.fitsInCpu?"Yes":"No"}`),console.log(`Limiting Factor: ${V.fit.limiting}`),V.limitedFit)console.log(`
+--- Memory-Limited Fit Analysis ---`),console.log(`Limited Total Required: ${(V.limitedFit.totalRequiredBytes/1024/1024/1024).toFixed(2)} GB`),console.log(`Fits in GPU (Limited): ${V.limitedFit.fitsInGpu?"Yes":"No"}`),console.log(`Fits in CPU (Limited): ${V.limitedFit.fitsInCpu?"Yes":"No"}`),console.log(`Limiting Factor (Limited): ${V.limitedFit.limiting}`)}}console.log(`
+=== Full Capabilities JSON ===`),console.log(JSON.stringify(_,null,2)),process.exit(0)}catch(X){console.error("Failed to get capabilities:",X.message),process.exit(1)}}async function F2({modelIds:$=[],defaultConfig:Z=null}={}){let X=[];if(console.log(`${C.name} v${C.version}`),console.log(`Generating STT model capabilities comparison...
+`),X.push(`${C.name} v${C.version}`),X.push(`## STT Model Capabilities Comparison
+`),!$||$.length===0)console.error("Error: No model IDs provided"),process.exit(1);try{let j=(U={},Y={})=>{let K=Array.isArray(U)?[...U]:{...U};return Object.entries(Y||{}).forEach(([L,B])=>{if(B&&typeof B==="object"&&!Array.isArray(B))K[L]=j(K[L]||{},B);else K[L]=B}),K},W=Z||{},{server:H,generators:Q=[],...J}=W,N=(U)=>j(JSON.parse(JSON.stringify(J)),U||{}),G=(U)=>{if(Array.isArray(Q)&&Q.length>0){let Y=Q.filter((K)=>K?.type==="ggml-stt");if(Y.length>0&&U){let K=Y.find((L)=>L.model?.repo_id===U);if(K)return N(K)}}return Object.keys(J).length>0?N({}):null},_=[];for(let U=0;U<$.length;U+=1){let Y=$[U],{repoId:K,filename:L}=R2(Y);console.log(`[${U+1}/${$.length}] Analyzing ${Y}...`);let B=G(K);B={...B||{},model:{...B?.model||{},repo_id:K,...L&&{filename:L}}};let E=await r("ggml-stt",null,{config:B,includeBreakdown:!0});_.push({modelId:Y,repoId:K,filename:L,capabilities:E,modelInfo:E.buttress?.selected||null,modelConfig:E.modelConfig||null})}let O=(U)=>U?(U/1024/1024).toFixed(1):"N/A",z=(U)=>U?"✅":"\uD83D\uDEAB";X.push("| Model | Size (MB) | Processing Buffer (MB) | Total Required (MB) | Fits GPU | Fits CPU |"),X.push("|-------|-----------|------------------------|---------------------|----------|----------|"),_.forEach(({modelId:U,modelInfo:Y})=>{let K=O(Y?.modelBytes),L=O(Y?.processingBytes||Y?.kvCacheBytes),B=O(Y?.fit?.totalRequiredBytes),E=z(Y?.fit?.fitsInGpu),M=z(Y?.fit?.fitsInCpu);X.push(`| ${U} | ${K} | ${L} | ${B} | ${E} | ${M} |`)}),X.push(`
+---`),X.push(`
+### System Information`);let V=null;if(process.platform!=="win32")try{V=o0("uname -a",{encoding:"utf8"}).trim()}catch{}if(V)X.push(`- **System:** ${V}`);else X.push(`- **Hostname:** ${y.hostname()}`),X.push(`- **OS:** ${y.type()} ${y.release()}`);if(X.push(`- **Platform:** ${process.platform}`),X.push(`- **CPU Cores:** ${y.cpus().length}`),X.push(`- **Total System Memory:** ${(y.totalmem()/1024/1024/1024).toFixed(2)} GB`),_.length>0){let Y=_[0].capabilities.buttress?.selected;if(Y){let K=Y.cpuTotalBytes>0?(Y.cpuUsableBytes/Y.cpuTotalBytes*100).toFixed(0):0;if(X.push(`- **Usable CPU Memory:** ${(Y.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${K}% of ${(Y.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),Y.hasGpu){let L=Y.devices.filter((B)=>B.type==="gpu");if(L.length>0){let B=L[0];X.push(`- **GPU Backend:** ${B.backend}`),X.push(`- **GPU Name:** ${B.deviceName}`),X.push(`- **GPU Total Memory:** ${(B.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let E=Y.gpuTotalBytes>0?(Y.gpuUsableBytes/Y.gpuTotalBytes*100).toFixed(0):0;X.push(`- **GPU Usable Memory:** ${(Y.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${E}% of ${(Y.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else X.push("- **GPU:** Not available")}}X.push(`
+### Command Used`);let R=process.argv.slice(2).join(" ");if(X.push(`\`\`\`bash
+${process.argv[0]} ${process.argv[1]} ${R}
+\`\`\``),X.push(`
+### Package Information`),X.push(`- **Name:** ${C.name}`),X.push(`- **Version:** ${C.version}`),C.description)X.push(`- **Description:** ${C.description}`);if(Z&&Object.keys(Z).length>0){X.push(`
+### Configuration`),X.push("<details>"),X.push("<summary>Click to expand TOML configuration</summary>"),X.push("\n```toml");try{let U=w2.stringify(Z);X.push(U)}catch(U){X.push("# Error serializing config"),X.push(JSON.stringify(Z,null,2))}X.push("```"),X.push("</details>")}let A=`ggml-stt-model-capabilities-${new Date().toISOString().replace(/[.:]/g,"-").split("T")[0]}.md`,w=B2.join(process.cwd(),A);z2.writeFileSync(w,X.join(`
+`),"utf8"),console.log(`
+STT model capabilities table saved to: ${w}`),process.exit(0)}catch(j){console.error("Failed to generate STT model table:",j.message),process.exit(1)}}async function x2({modelId:$=null,defaultConfig:Z=null}={}){if(console.log(`${C.name} v${C.version}`),console.log("Testing capabilities for backend: ggml-stt"),$)console.log(`Model: ${$}`);console.log("--------------------------------");try{let{repoId:X,filename:j}=R2($),W=Z||{},{server:H,generators:Q=[],...J}=W,N=(q={},A={})=>{let w=Array.isArray(q)?[...q]:{...q};return Object.entries(A||{}).forEach(([U,Y])=>{if(Y&&typeof Y==="object"&&!Array.isArray(Y))w[U]=N(w[U]||{},Y);else w[U]=Y}),w},G=(q)=>N(JSON.parse(JSON.stringify(J)),q||{}),O=((q)=>{if(Array.isArray(Q)&&Q.length>0){let A=Q.filter((w)=>w?.type==="ggml-stt");if(A.length>0){if(q){let w=A.find((U)=>U.model?.repo_id===q);if(w)return G(w)}}}if(Object.keys(J).length>0)return G({});return null})(X);if(X)O={...O||{},model:{...J.runtime,...O?.model||{},repo_id:X,...j&&{filename:j}}};let z=await r("ggml-stt",null,{config:O,includeBreakdown:!0}),V=z.buttress?.selected||null,R=z.modelConfig||null;if(X||R?.repoId){console.log(`
+=== Model Information ===`);let q=X||R?.repoId;if(console.log(`Repository ID: ${q}`),j)console.log(`Filename: ${j}`);if(V?.modelBytes)console.log(`Model Size: ${(V.modelBytes/1024/1024).toFixed(1)} MB`);let A=V?.processingBytes||V?.kvCacheBytes;if(A)console.log(`Processing Buffer: ${(A/1024/1024).toFixed(1)} MB`);if(V?.modelBytes&&A)console.log(`Total Required Memory: ${((V.modelBytes+A)/1024/1024).toFixed(1)} MB`);else if(z.buttress?.selected?.fit){let{totalRequiredBytes:w}=z.buttress.selected.fit;console.log(`Total Required Memory: ${(w/1024/1024).toFixed(1)} MB`)}}if(z.buttress?.selected){let{selected:q}=z.buttress;console.log(`
+=== Hardware Information ===`);let A=null;if(process.platform!=="win32")try{A=o0("uname -a",{encoding:"utf8"}).trim()}catch{}if(A)console.log(`System: ${A}`);else console.log(`Hostname: ${y.hostname()}`),console.log(`OS: ${y.type()} ${y.release()}`);console.log(`Platform: ${q.platform}`),console.log(`CPU Cores: ${y.cpus().length}`),console.log(`Total System Memory: ${(y.totalmem()/1024/1024/1024).toFixed(2)} GB`);let w=q.cpuTotalBytes>0?(q.cpuUsableBytes/q.cpuTotalBytes*100).toFixed(0):0;if(console.log(`Usable CPU Memory: ${(q.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${w}% of ${(q.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),q.hasGpu)console.log(`
+--- GPU Details ---`),q.devices.filter((Y)=>Y.type==="gpu").forEach((Y)=>{console.log(`GPU Backend: ${Y.backend}`),console.log(`GPU Name: ${Y.deviceName}`),console.log(`GPU Total Memory: ${(Y.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let K=q.gpuTotalBytes>0?(q.gpuUsableBytes/q.gpuTotalBytes*100).toFixed(0):0;if(console.log(`GPU Usable Memory: ${(q.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${K}% of ${(q.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),Y.metadata){if(Y.metadata.hasBFloat16)console.log("Supports BFloat16: Yes");if(Y.metadata.hasUnifiedMemory)console.log("Unified Memory: Yes")}});else console.log("GPU: Not available");if(console.log(`
+Backend Variant: ${q.variant}`),console.log(`Performance Score: ${q.score}`),q.fit)console.log(`
+--- Model Fit Analysis ---`),console.log(`Fits in GPU: ${q.fit.fitsInGpu?"Yes":"No"}`),console.log(`Fits in CPU: ${q.fit.fitsInCpu?"Yes":"No"}`),console.log(`Limiting Factor: ${q.fit.limiting}`)}console.log(`
+=== Full Capabilities JSON ===`),console.log(JSON.stringify(z,null,2)),process.exit(0)}catch(X){console.error("Failed to get capabilities:",X.message),process.exit(1)}}var R2=($)=>{if(!$)return{repoId:null,filename:null};let[Z,X]=$.split(":");return{repoId:Z,filename:X||null}};var S2=S(()=>{B0();d1();A2()});var e0={};F3(e0,{testGgmlSttCapabilities:()=>x2,testGgmlLlmCapabilities:()=>M2,status:()=>c6,startModelDownload:()=>t0,startGenerator:()=>u6,showSttModelsTable:()=>F2,showModelsTable:()=>E2,globalDownloadManager:()=>l1,ggmlStt:()=>f6,ggmlLlm:()=>p6,getModelIdentifier:()=>g6,getCapabilities:()=>r,generatorRegistry:()=>$0,finalizeGenerator:()=>m6});async function u6($,Z){let j={"ggml-llm":{create:a4,getId:b1},"ggml-stt":{create:q2,getId:g1}}[$];if(!j)throw Error(`Unsupported backend type: ${$}`);let W=j.getId(Z);if(!W)throw Error("Buttress generator config missing repo identifier");let H=`${$}:${W}`,Q=$0.get(H);if(Q)return Q.refCount+=1,Q.instance.resetFinalized?.(),{id:Q.id,info:Q.instance.info};let J=await j.create(H,Z,{globalDownloadManager:l1}),N={id:H,type:J.type,instance:J,refCount:1};return $0.set(H,N),{id:H,info:J.info}}async function m6($){let Z=$0.get($);if(!Z)return!1;if(Z.refCount-=1,Z.refCount<=0){if(await Z.instance.finalize(),!(Z.instance.hasPendingReleases?.()??!1))$0.delete($)}return!0}function g6($,Z){if($==="ggml-llm")return b1(Z);if($==="ggml-stt")return g1(Z);return null}async function t0($,Z,X={}){let W={"ggml-llm":o4,"ggml-stt":O2}[$];if(!W)return{started:!1,localPath:null,repoId:null,error:`Unknown backend type: ${$}`};return W(Z,l1,X)}var $0,l1,y6=($)=>{let Z=$0.get($);if(!Z)throw Error(`Unknown generator id "${$}"`);return Z},V0=($,Z)=>{let X=y6($);if(X.type!==Z)throw Error(`Generator "${$}" does not support ${Z} backend`);return X.instance},p6,f6,c6;var $1=S(()=>{v1();c1();d1();m0();S2();$0=new Map,l1={downloads:new Map,getDownload($){return this.downloads.get($)||null},setDownload($,Z){this.downloads.set($,Z)},deleteDownload($){this.downloads.delete($)},isDownloading($){return this.downloads.has($)},getActiveDownloads(){return Array.from(this.downloads.entries()).map(([$,Z])=>({localPath:$,promise:Z}))}};p6={async initContext($,Z){return V0($,"ggml-llm").initContext(Z)},async completion($,Z){return V0($,"ggml-llm").completion(Z)},async tokenize($,Z){return V0($,"ggml-llm").tokenize(Z)},async detokenize($,Z){return V0($,"ggml-llm").detokenize(Z)},async applyChatTemplate($,Z){return V0($,"ggml-llm").applyChatTemplate(Z)},async releaseContext($,Z){let X=$0.get($);if(!X)return{released:!0,alreadyReleased:!0};if(X.type!=="ggml-llm")throw Error(`Generator "${$}" does not support ggml-llm backend`);return X.instance.releaseContext(Z)}},f6={async initContext($,Z){return V0($,"ggml-stt").initContext(Z)},async transcribe($,Z){return V0($,"ggml-stt").transcribe(Z)},async transcribeData($,Z){return V0($,"ggml-stt").transcribeData(Z)},async releaseContext($,Z){let X=$0.get($);if(!X)return{released:!0,alreadyReleased:!0};if(X.type!=="ggml-stt")throw Error(`Generator "${$}" does not support ggml-stt backend`);return X.instance.releaseContext(Z)}};c6={getFullStatus:()=>x4($0),getGgmlLlmStatus:()=>F1($0),getGgmlSttStatus:()=>x1($0),subscribeToStatus:M1,subscribeToStatusWithId:F4,llmStatusTracker:t,sttStatusTracker:J0,statusEmitter:o}});import{node as d6}from"@elysiajs/node";import{Elysia as l6}from"elysia";var n6,U0=($)=>new l6({adapter:n6?d6():void 0,...$});var F0=S(()=>{n6=typeof process<"u"&&process.versions&&process.versions.node});import{t as j0}from"elysia";var i6,s6=({store:{serverInfo:$}})=>({id:$.id,name:$.name,version:$.version,generators:$.generators,authentication:$.authentication}),n1=($)=>{let Z=U0(),X=$.autodiscover.http?.path??"/buttress/info";return Z.get(X,s6,{response:i6}),Z};var P2=S(()=>{F0();i6=j0.Object({id:j0.String(),name:j0.String(),version:j0.String(),generators:j0.Array(j0.Object({type:j0.String()})),authentication:j0.Object({required:j0.Boolean(),type:j0.Literal("device-group")})})});import{t as W0,file as r6}from"elysia";import{writeFile as T2}from"node:fs/promises";import i1 from"node:path";var a6,s1;var k2=S(()=>{F0();a6=typeof process<"u"&&process.versions!=null&&process.versions.node!=null,s1=U0().post("/buttress/upload",async({body:{file:$},store:{config:Z}})=>{let X=`${Date.now()}-${$.name.replace(/[^\dA-Za-z]/g,"_")}`,j=i1.join(Z.server.temp_file_dir,X);try{if(a6)await T2(j,await $.stream());else await T2(j,await $.arrayBuffer());return{ok:!0,filename:X}}catch(W){return{ok:!1,error:String(W)}}},{body:W0.Object({file:W0.File()}),response:W0.Object({ok:W0.Boolean(),filename:W0.Optional(W0.String()),error:W0.Optional(W0.String())})}).get("/buttress/download/:filename",async({params:{filename:$},store:{config:Z},status:X})=>{let j=i1.join(Z.server.temp_file_dir,$);if(i1.relative(Z.server.temp_file_dir,j).includes(".."))return X(400),"Invalid file path";return r6(j)},{params:W0.Object({filename:W0.String()})})});import b2 from"node:path";import o6 from"node:fs/promises";import{fileURLToPath as t6}from"node:url";var e6,$$=($)=>{let{status:Z}=$;if(Z?.getFullStatus)return Z.getFullStatus();return{timestamp:new Date().toISOString(),ggmlLlm:{generators:[],history:{}},ggmlStt:{generators:[],history:{}}}},D2=async()=>{try{let $=b2.join(e6,"..","..","public","status.html"),Z=await o6.readFile($,"utf-8");return new Response(Z,{headers:{"Content-Type":"text/html; charset=utf-8"}})}catch($){return console.error("[Status] Failed to serve status page:",$),new Response("Status page not found",{status:404,headers:{"Content-Type":"text/plain"}})}},r1;var v2=S(()=>{F0();e6=b2.dirname(t6(import.meta.url)),r1=U0().get("/status",D2).get("/status/",D2).get("/buttress/status",({store:{backend:$}})=>$$($))});import{t as P,sse as Z1}from"elysia";import{cors as Z$}from"@elysiajs/cors";async function j$($,Z,X){let W=(Z.generators||[]).filter((q)=>q.type==="ggml-llm");if(W.length===0)throw Error('No ggml-llm generator configured. Add a [[generators]] with type = "ggml-llm" to your config.');let H=W[0],Q=X||H.model?.repo_id;if(X){let q=W.find((A)=>A.model?.repo_id===X);if(q)H=q}else Q=H.model?.repo_id;let J=Q,N=h2.get(J);if(N?.initialized)return N;let{generators:G,server:_,...O}=Z.global||{},z={...O,...H,model:{...H.model,repo_id:Q}};console.log(`[OpenAI] Creating generator for ${J}`);let{id:V}=await $.startGenerator("ggml-llm",z),R={id:V,config:z,repoId:Q,initialized:!1};return h2.set(J,R),await $.ggmlLlm.initContext(V,{}),R.initialized=!0,console.log(`[OpenAI] Generator ready: ${J}`),R}async function W$($,Z,X,j){let W=$.getReader(),H="",Q=null,J=null,N="stop",G={prompt_tokens:0,completion_tokens:0,total_tokens:0};try{let O=!1;while(!O){let z=await W.read();if({done:O}=z,O)break;let{event:V,data:R}=z.value;if(V==="token"){if(R.content!=null)H+=R.content;else if(R.token!=null)H+=R.token}else if(V==="result"){if(R.text)H=R.text;else if(R.content)H=R.content;if(R.reasoning_content)Q=R.reasoning_content;if(R.tool_calls?.length>0)J=R.tool_calls.map((q,A)=>({id:q.id||`call_${Z}_${A}`,type:"function",function:{name:q.function?.name||"",arguments:q.function?.arguments||""}})),N="tool_calls";else N=R.interrupted?"length":"stop";G={prompt_tokens:R.prompt_tokens||R.promptTokens||0,completion_tokens:R.tokens_predicted||R.tokensPredicted||0,total_tokens:(R.prompt_tokens||R.promptTokens||0)+(R.tokens_predicted||R.tokensPredicted||0)}}else if(V==="error")throw Error(R.message)}}finally{W.cancel().catch(()=>{})}let _={role:"assistant",content:H||null};if(Q)_.reasoning_content=Q;if(J)_.tool_calls=J;return{id:Z,object:"chat.completion",created:X,model:j,choices:[{index:0,message:_,finish_reason:N}],usage:G}}function X1({global:$}){let Z=U0({prefix:"/oai-compat"});return Z.use(Z$({origin:$?.openai_compat?.cors_allowed_origins??!1,methods:["GET","POST","OPTIONS"],allowedHeaders:["Content-Type","Authorization"],maxAge:86400,preflight:!0})),Z.get("/v1/models",({store:X})=>{let{config:j}=X,Q=(j.generators||[]).filter((J)=>J.type==="ggml-llm").map((J)=>{return{id:J.model?.repo_id||"ggml-llm",object:"model",created:Math.floor(Date.now()/1000),owned_by:"local"}});if(Q.length===0)Q.push({id:"ggml-llm",object:"model",created:Math.floor(Date.now()/1000),owned_by:"local"});return{object:"list",data:Q}}),Z.post("/v1/chat/completions",async function*({body:j,set:W,store:H}){let{config:Q,backend:J}=H,{messages:N=[],stream:G=!1,model:_,tools:O,temperature:z,stop:V,top_p:R,max_tokens:q,presence_penalty:A,frequency_penalty:w,tool_choice:U,stream_options:Y}=j;if(!N||N.length===0)return W.status=400,{error:{message:"messages is required and must not be empty",type:"invalid_request_error"}};try{let K=await j$(J,Q,_),L=X$(),B=Math.floor(Date.now()/1000),E=K.repoId||"ggml-llm",M={reasoning_format:"auto",messages:N,jinja:!0,add_generation_prompt:!0};if(z!=null)M.temperature=z;if(R!=null)M.top_p=R;if(q!=null)M.n_predict=q;if(V!=null)M.stop=Array.isArray(V)?V:[V];if(A!=null)M.presence_penalty=A;if(w!=null)M.frequency_penalty=w;if(O!=null)M.tools=O;if(U!=null)M.tool_choice=U;let F=await J.ggmlLlm.completion(K.id,{options:M});if(!G)return await W$(F,L,B,E);let x=Y?.include_usage===!0,k=F.getReader(),D="",m="",p=new Map,n=new Map;try{let i=!1;while(!i){let v=await k.read();if({done:i}=v,i)break;let{event:c,data:T}=v.value;if(c==="token"){let f={};if(T.content!=null){let h=T.content;if(h.length>D.length)f.content=h.slice(D.length),D=h}if(T.reasoning_content!=null){let h=T.reasoning_content;if(h.length>m.length)f.reasoning_content=h.slice(m.length),m=h}if(T.tool_calls?.length>0){let h=[];if(T.tool_calls.forEach((s,I)=>{let a={index:I};if(!n.has(I))n.set(I,s.id||`call_${L}_${I}`),a.id=n.get(I),a.type="function";let q0=s.function?.arguments||"",T0=p.get(I)||"",G0={};if(!p.has(I)&&s.function?.name)G0.name=s.function.name;if(q0.length>T0.length)G0.arguments=q0.slice(T0.length),p.set(I,q0);if(Object.keys(G0).length>0)a.function=G0,h.push(a);else if(a.id)a.function={name:s.function?.name||"",arguments:""},h.push(a)}),h.length>0)f.tool_calls=h}if(Object.keys(f).length>0)yield Z1({data:JSON.stringify({id:L,object:"chat.completion.chunk",created:B,model:E,choices:[{index:0,delta:f,finish_reason:null}]})})}else if(c==="result"){let f={id:L,object:"chat.completion.chunk",created:B,model:E,choices:[{index:0,delta:{},finish_reason:T.interrupted?"length":"stop"}]};if(x)f.usage={prompt_tokens:T.prompt_tokens||T.promptTokens||0,completion_tokens:T.tokens_predicted||T.tokensPredicted||0,total_tokens:(T.prompt_tokens||T.promptTokens||0)+(T.tokens_predicted||T.tokensPredicted||0)};yield Z1({data:JSON.stringify(f)})}else if(c==="error")yield Z1({data:JSON.stringify({error:{message:T.message,type:"server_error"}})})}yield Z1({data:"[DONE]"})}finally{k.cancel().catch(()=>{})}}catch(K){return console.error("[OpenAI] Chat completion error:",K),W.status=500,{error:{message:K.message||"Internal server error",type:"server_error"}}}},{body:P.Object({model:P.Optional(P.String()),messages:P.Array(P.Any()),stream:P.Optional(P.Boolean()),temperature:P.Optional(P.Number()),top_p:P.Optional(P.Number()),max_tokens:P.Optional(P.Number()),stop:P.Optional(P.Union([P.String(),P.Array(P.String())])),presence_penalty:P.Optional(P.Number()),frequency_penalty:P.Optional(P.Number()),tools:P.Optional(P.Array(P.Any())),tool_choice:P.Optional(P.Any()),stream_options:P.Optional(P.Object({include_usage:P.Optional(P.Boolean())}))})}),Z}var X$=()=>`chatcmpl-${Date.now()}-${Math.random().toString(36).slice(2,9)}`,h2;var C2=S(()=>{F0();h2=new Map});var I2=S(()=>{P2();k2();v2();C2()});import Q$ from"node:os";import J$ from"node:path";import N$ from"node-machine-id";import H$ from"bytes";import V$ from"ms";var x0=($={},Z={})=>{let X=Array.isArray($)?[...$]:{...$};return Object.entries(Z||{}).forEach(([j,W])=>{if(W&&typeof W==="object"&&!Array.isArray(W))X[j]=x0(X[j]||{},W);else X[j]=W}),X},A0=($)=>{if(!$)return null;if(typeof $==="object")return JSON.parse(JSON.stringify($));return null},a1=($,Z)=>{let X=A0($)||{},j=A0(Z)||{};return x0(X,j)},y2=($,Z)=>x0(JSON.parse(JSON.stringify($.global)),Z||{}),o1=($,Z,X,j)=>{if($.generators.length>0){let W=$.generators.filter((H)=>H?.type===X);if(W.length>0){if(j){let H=W.find((Q)=>Z.getModelIdentifier(X,Q)===j);if(H)return y2($,H)}}}if(Object.keys($.global).length>0)return y2($,{});return null},u2,U$=($)=>{if(!$)return null;if($===!0)return{...u2};return x0(u2,$)},g2=($,Z)=>{if(!$.generators||$.generators.length===0)return Z.map((j)=>({type:j}));let X=new Set;if($.generators.forEach((j)=>{if(j.type)X.add(j.type)}),X.size===0)return Z.map((j)=>({type:j}));return Array.from(X).map((j)=>({type:j}))},m2=($,Z,X)=>{if($===void 0)return X;if(typeof $==="number")return $;return Z($)??X},G$=2080,p2=60000,f2=52428800,t1=($)=>{let Z=N$.machineIdSync(),X={server:{id:`buttress-${Z}`,name:`Buttress Server (${Z.slice(-8)})`,port:G$,temp_file_dir:J$.join(Q$.tmpdir(),".buttress"),session_timeout:p2,max_body_size:f2},autodiscover:!1},j=x0(X,A0($)||{}),W=Array.isArray(j.generators)?j.generators:[],{server:H,generators:Q,autodiscover:J,...N}=j;return{autodiscover:U$(J),server:{id:H.id,name:H.name,port:H.port,log_level:H.log_level,temp_file_dir:H.temp_file_dir,max_body_size:m2(H.max_body_size,H$.parse,f2),session_timeout:m2(H.session_timeout,V$,p2)},global:N,generators:W}};var S0=S(()=>{u2={udp:{port:8089,announcements:{enabled:!0,interval:5000},requests:{enabled:!0,responseDelay:100}},http:{enabled:!0,path:"/buttress/info",cors:!0}}});import{z as Z0}from"zod";var c2,d2;var l2=S(()=>{S0();c2={getCapabilities:Z0.tuple([Z0.object({type:Z0.string().optional().default("ggml-llm"),config:Z0.any().optional(),currentClientCapabilities:Z0.any().optional(),options:Z0.any().optional()}).nullable().optional()]),startGenerator:Z0.tuple([Z0.string(),Z0.any().optional()]),finalizeGenerator:Z0.tuple([Z0.string()])},d2={async getCapabilities({backend:$,config:Z},X=null){console.log("[Server] Get Capabilities:",X);let j=X||{type:"ggml-llm"},{type:W="ggml-llm",config:H,currentClientCapabilities:Q=null,options:J={}}=j,N=A0(H),G=$.getModelIdentifier(W,N),_=o1(Z,$,W,G),O=a1(_,H);if(Object.keys(O).length===0)throw Error("Buttress server missing generator configuration");if(O.backend=O.backend||{},!O.backend.type)O.backend.type=W;return $.getCapabilities(W,Q,{...J,config:O})},async startGenerator({backend:$,config:Z,session:X},j,W){console.log("[Server] Start Generator:",j,W);let H=A0(W),Q=$.getModelIdentifier(j,H),J=o1(Z,$,j,Q),N=a1(J,W);if(Object.keys(N).length===0)throw Error("Buttress server missing generator configuration");if(N.backend=N.backend||{},!N.backend.type)N.backend.type=j;let G=await $.startGenerator(j,N);return X.generators.add(G.id),G},async finalizeGenerator({backend:$,session:Z},X){return console.log("[Server] Finalize Generator:",X),Z.generators.delete(X),$.finalizeGenerator(X)}}});import{z as g}from"zod";import{ReadableStream as K$}from"node:stream/web";var n2,i2;var s2=S(()=>{n2={initContext:g.tuple([g.string(),g.any().optional()]),completion:g.tuple([g.string(),g.any().optional()]),tokenize:g.tuple([g.string(),g.any()]),detokenize:g.tuple([g.string(),g.any()]),applyChatTemplate:g.tuple([g.string(),g.any()]),releaseContext:g.tuple([g.string(),g.boolean().optional()])},i2={initContext({backend:$},Z,X){return new K$({async start(j){try{let W=await $.ggmlLlm.initContext(Z,{...X,onProgress:(J)=>{j.enqueue({event:"progress",data:{progress:J}})}});await new Promise((J)=>setTimeout(J,1000));let{download:H,...Q}=W||{};j.enqueue({event:"result",data:{result:Q}}),j.close()}catch(W){j.error(W)}}})},completion({backend:$},Z,X){return console.log("[Server] Completion:",{id:Z,property:X}),$.ggmlLlm.completion(Z,X)},async tokenize({backend:$},Z,X){return console.log("[Server] Tokenize:",{id:Z,property:X}),$.ggmlLlm.tokenize(Z,X)},async detokenize({backend:$},Z,X){return console.log("[Server] Detokenize:",{id:Z,property:X}),$.ggmlLlm.detokenize(Z,X)},async applyChatTemplate({backend:$},Z,X){return console.log("[Server] Apply Chat Template:",{id:Z,property:X}),$.ggmlLlm.applyChatTemplate(Z,X)},async releaseContext({backend:$},Z,X){return console.log("[Server] Release Context:",{id:Z,force:X}),$.ggmlLlm.releaseContext(Z,{force:X})}}});import{z as l}from"zod";import{ReadableStream as _$}from"node:stream/web";import q$ from"node:path";var r2,a2;var o2=S(()=>{r2={initContext:l.tuple([l.string(),l.any().optional()]),transcribe:l.tuple([l.string(),l.string(),l.any().optional()]),transcribeData:l.tuple([l.string(),l.union([l.instanceof(Buffer),l.instanceof(Uint8Array)]),l.any().optional()]),releaseContext:l.tuple([l.string(),l.boolean().optional()])},a2={initContext({backend:$},Z,X){return new _$({async start(j){try{let W=await $.ggmlStt.initContext(Z,{...X,onProgress:(J)=>{j.enqueue({event:"progress",data:{progress:J}})}});await new Promise((J)=>setTimeout(J,1000));let{download:H,...Q}=W||{};j.enqueue({event:"result",data:{result:Q}}),j.close()}catch(W){j.error(W)}}})},async transcribe({backend:$,config:{server:Z}},X,j,W){return console.log("[Server] Transcribe:",{id:X,audioPath:j,options:W}),$.ggmlStt.transcribe(X,{audioPath:q$.join(Z.temp_file_dir,j),options:W})},async transcribeData({backend:$},Z,X,j){return console.log("[Server] Transcribe Data:",{id:Z,audioDataLength:X?.length||0,options:j}),$.ggmlStt.transcribeData(Z,{audioData:X,options:j})},async releaseContext({backend:$},Z,X){return console.log("[Server] Release STT Context:",{id:Z,force:X}),$.ggmlStt.releaseContext(Z,{force:X})}}});var O$,t2,e2;var $3=S(()=>{l2();s2();o2();O$={common:d2,ggmlLlm:i2,ggmlStt:a2},t2={common:c2,ggmlLlm:n2,ggmlStt:r2},e2=O$});import{Buffer as j1}from"node:buffer";var Z3=($)=>{try{return JSON.parse($,(Z,X)=>{if(!X)return X;if(X?.type==="Buffer"&&X?.data)return j1.from(X.data,"base64");if(X?.type==="Uint8Array"&&X?.data){let j=j1.from(X.data,"base64");return j.buffer.slice(j.byteOffset,j.byteOffset+j.byteLength)}if(X?.type==="Error"&&X?.name&&X?.message)return Error(X.name,X.message);return X})}catch{return $}},W1=($)=>{try{return JSON.stringify($,(Z,X)=>{if(X instanceof Error)return{type:"Error",name:X.name,message:X.message};if(X instanceof j1)return{type:"Buffer",data:X.toString("base64")};if(X instanceof Uint8Array)return{type:"Uint8Array",data:j1.from(X).toString("base64")};return X})}catch{return $}};var X3=()=>{};var e1="1.0",j3=8089;import L$ from"node:dgram";class $4{name="udp";socket=null;announcementTimer=null;config;getServerInfo;port;constructor($,Z){this.config=$,this.getServerInfo=Z,this.port=$.port??j3}async start(){if(this.socket=L$.createSocket({type:"udp4",reuseAddr:!0}),this.socket.on("message",($,Z)=>{this.handleMessage($,Z)}),this.socket.on("error",($)=>{console.error("[Autodiscover UDP] Socket error:",$.message)}),await new Promise(($,Z)=>{this.socket.bind(this.port,()=>{this.socket.setBroadcast(!0),console.log(`[Autodiscover UDP] Listening on port ${this.port}`),$()}),this.socket.once("error",Z)}),this.config.announcements.enabled){let $=this.config.announcements.interval??5000;this.announcementTimer=setInterval(()=>{this.sendAnnouncement()},$),this.sendAnnouncement()}}async stop(){if(this.announcementTimer)clearInterval(this.announcementTimer),this.announcementTimer=null;if(this.socket)await new Promise(($)=>{this.socket.close(()=>$())}),this.socket=null}handleMessage($,Z){try{let X=JSON.parse($.toString());if(X.t==="QUERY"&&this.config.requests.enabled){let j=X.d,W=this.config.requests.responseDelay??0,H=W>0?Math.random()*W:0;setTimeout(()=>{this.sendResponse(j.id,Z)},H)}}catch{}}sendAnnouncement(){if(!this.socket)return;let $=this.getServerInfo(),Z={t:"ANNOUNCE",v:e1,d:{info:$}},X=Buffer.from(JSON.stringify(Z));this.socket.send(X,0,X.length,this.port,"255.255.255.255",(j)=>{if(j)console.error("[Autodiscover UDP] Announcement error:",j.message)})}sendResponse($,Z){if(!this.socket)return;let X=this.getServerInfo(),j={t:"RESPONSE",v:e1,d:{request_id:$,info:X}},W=Buffer.from(JSON.stringify(j));this.socket.send(W,0,W.length,Z.port,Z.address,(H)=>{if(H)console.error("[Autodiscover UDP] Response error:",H.message)})}}var W3=()=>{};class Z4{config;getServerInfo;transports=[];started=!1;constructor($,Z){this.config=$;this.getServerInfo=Z;if($.udp?.announcements?.enabled||$.udp?.requests?.enabled)this.transports.push(new $4($.udp,Z))}async start(){if(this.started)return;(await Promise.allSettled(this.transports.map((Z)=>Z.start()))).forEach((Z,X)=>{if(Z.status==="rejected")console.error(`[Autodiscover] Failed to start ${this.transports[X].name}:`,Z.reason)}),this.started=!0}async stop(){if(!this.started)return;await Promise.allSettled(this.transports.map(($)=>$.stop())),this.started=!1}}var Y3=S(()=>{W3()});import A$ from"node:os";var Y1=()=>{let $=A$.networkInterfaces();return Object.values($).flat().find((X)=>X?.family==="IPv4"&&!X?.internal)?.address||null};var X4=()=>{};import u from"node:os";import Q3 from"node:fs";import J3 from"node:path";import{execSync as Q1}from"node:child_process";import N3 from"@iarna/toml";async function j4({modelIds:$=[],defaultConfig:Z=null}={}){let X=[];if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.34"}`),console.log(`Generating model capabilities comparison...
+`),X.push(`${"@fugood/buttress-server"} v${"2.23.0-beta.34"}`),X.push(`## Model Capabilities Comparison
+`),!$||$.length===0)console.error("Error: No model IDs provided"),process.exit(1);try{let j=(U={},Y={})=>{let K=Array.isArray(U)?[...U]:{...U};return Object.entries(Y||{}).forEach(([L,B])=>{if(B&&typeof B==="object"&&!Array.isArray(B))K[L]=j(K[L]||{},B);else K[L]=B}),K},W=Z||{},{server:H,generators:Q=[],...J}=W,N=(U)=>j(JSON.parse(JSON.stringify(J)),U||{}),G=(U)=>{if(Array.isArray(Q)&&Q.length>0){let Y=Q.filter((K)=>K?.type==="ggml-llm");if(Y.length>0&&U){let K=Y.find((L)=>L.model?.repo_id===U);if(K)return N(K)}}return Object.keys(J).length>0?N({}):null},_=[];for(let U=0;U<$.length;U+=1){let Y=$[U];console.log(`[${U+1}/${$.length}] Analyzing ${Y}...`);let K=G(Y);K={...K||{},model:{...J.runtime,...K?.model||{},repo_id:Y}};let L=await r("ggml-llm",null,{config:K,includeBreakdown:!0});_.push({modelId:Y,capabilities:L,modelInfo:L.buttress?.selected||null,modelConfig:L.modelConfig||null})}let O=(U)=>U?(U/1024/1024/1024).toFixed(2):"N/A",z=(U)=>U?"✅":"\uD83D\uDEAB";X.push("| Model ID | Size (GB) | Context Size | KV Cache Size (GB) | Total Required Memory (GB) | Fits GPU (Full) | Fits CPU (Full) |"),X.push("|----------|-----------|--------------|--------------------|----------------------------|-----------------|-----------------|"),_.forEach(({modelId:U,modelInfo:Y,modelConfig:K})=>{let L=O(Y?.modelBytes),B=K?.nCtx||Y?.kvInfo?.nCtxTrain||"N/A",E=N0(Y),M=Number(B),F=Y?.kvCacheBytes||(E&&Number.isFinite(M)&&M>0?E(M):E&&E(Y?.kvInfo?.nCtxTrain||0))||null,x=O(F),k=O(Y?.modelBytes&&F?Y.modelBytes+F:Y?.fit?.totalRequiredBytes),D=z(Y?.fit?.fitsInGpu),m=z(Y?.fit?.fitsInCpu);X.push(`| ${U} | ${L} | ${B} | ${x} | ${k} | ${D} | ${m} |`);let p=Y?.memoryLimitedCtx!=null||Y?.limitedFit!=null,n=!Y?.fit?.fitsInGpu||!Y?.fit?.fitsInCpu;if(p&&n){let i=Y?.memoryLimitedCtx||B,v=Number(i),c=Y?.limitedKvCacheBytes||E&&Number.isFinite(v)&&v>0&&E(v)||null,T=O(c),f=O(Y?.modelBytes&&c?Y.modelBytes+c:Y?.limitedFit?.totalRequiredBytes),h=z(Y?.limitedFit?.fitsInGpu),s=z(Y?.limitedFit?.fitsInCpu);if(i!==B||T!==x||f!==k)X.push(`| ↳ Limited | ${L} | ${i} | ${T} | ${f} | ${h} | ${s} |`)}}),X.push(`
+---`),X.push(`
+### System Information`);let V=null;if(process.platform!=="win32")try{V=Q1("uname -a",{encoding:"utf8"}).trim()}catch{}if(V)X.push(`- **System:** ${V}`);else X.push(`- **Hostname:** ${u.hostname()}`),X.push(`- **OS:** ${u.type()} ${u.release()}`);if(X.push(`- **Platform:** ${process.platform}`),X.push(`- **CPU Cores:** ${u.cpus().length}`),X.push(`- **Total System Memory:** ${(u.totalmem()/1024/1024/1024).toFixed(2)} GB`),_.length>0){let Y=_[0].capabilities.buttress?.selected;if(Y){let K=Y.cpuTotalBytes>0?(Y.cpuUsableBytes/Y.cpuTotalBytes*100).toFixed(0):0;if(X.push(`- **Usable CPU Memory:** ${(Y.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${K}% of ${(Y.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),Y.hasGpu){let L=Y.devices.filter((B)=>B.type==="gpu");if(L.length>0){let B=L[0];X.push(`- **GPU Backend:** ${B.backend}`),X.push(`- **GPU Name:** ${B.deviceName}`),X.push(`- **GPU Total Memory:** ${(B.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let E=Y.gpuTotalBytes>0?(Y.gpuUsableBytes/Y.gpuTotalBytes*100).toFixed(0):0;X.push(`- **GPU Usable Memory:** ${(Y.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${E}% of ${(Y.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else X.push("- **GPU:** Not available")}}X.push(`
+### Command Used`);let R=process.argv.slice(2).join(" ");if(X.push(`\`\`\`bash
+${process.argv[0]} ${process.argv[1]} ${R}
+\`\`\``),X.push(`
+### Package Information`),X.push(`- **Name:** ${"@fugood/buttress-server"}`),X.push(`- **Version:** ${"2.23.0-beta.34"}`),Z&&Object.keys(Z).length>0){X.push(`
+### Configuration`),X.push("<details>"),X.push("<summary>Click to expand TOML configuration</summary>"),X.push("\n```toml");try{let U=N3.stringify(Z);X.push(U)}catch(U){X.push("# Error serializing config"),X.push(JSON.stringify(Z,null,2))}X.push("```"),X.push("</details>")}let A=`ggml-llm-model-capabilities-${new Date().toISOString().replace(/[.:]/g,"-").split("T")[0]}.md`,w=J3.join(process.cwd(),A);Q3.writeFileSync(w,X.join(`
+`),"utf8"),console.log(`
+Model capabilities table saved to: ${w}`),process.exit(0)}catch(j){console.error("Failed to generate model table:",j.message),process.exit(1)}}async function V3({modelId:$=null,defaultConfig:Z=null}={}){if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.34"}`),console.log("Testing capabilities for backend: ggml-llm"),$)console.log(`Model: ${$}`);console.log("--------------------------------");try{let X=Z||{},{server:j,generators:W=[],...H}=X,Q=(V={},R={})=>{let q=Array.isArray(V)?[...V]:{...V};return Object.entries(R||{}).forEach(([A,w])=>{if(w&&typeof w==="object"&&!Array.isArray(w))q[A]=Q(q[A]||{},w);else q[A]=w}),q},J=(V)=>Q(JSON.parse(JSON.stringify(H)),V||{}),G=((V)=>{if(Array.isArray(W)&&W.length>0){let R=W.filter((q)=>q?.type==="ggml-llm");if(R.length>0){if(V){let q=R.find((A)=>A.model?.repo_id===V);if(q)return J(q)}}}if(Object.keys(H).length>0)return J({});return null})($);if($)G={...G||{},model:{...G?.model||{},repo_id:$}};let _=await r("ggml-llm",null,{config:G,includeBreakdown:!0}),O=_.buttress?.selected||null,z=_.modelConfig||null;if($||z?.repoId){console.log(`
+=== Model Information ===`);let V=$||z?.repoId;if(console.log(`Repository ID: ${V}`),z?.quantization)console.log(`Quantization: ${z.quantization}`);if(z?.nCtx)console.log(`Context Length: ${z.nCtx}`);if(O?.quantization){let{fileType:A}=O.quantization;if(A!=null)console.log(`Model File Type (GGUF): ${A}`)}let R=z?.cache_type_k||"f16",q=z?.cache_type_v||"f16";if(console.log(`KV Cache Type: K=${R}, V=${q}`),O?.modelBytes&&O?.kvCacheBytes){if(console.log(`Model Size: ${(O.modelBytes/1024/1024/1024).toFixed(2)} GB`),O.kvInfo)console.log(`KV Cache Size: ${(O.kvCacheBytes/1024/1024/1024).toFixed(2)} GB (KV info: ${JSON.stringify(O.kvInfo)})`);else console.log(`KV Cache Size: ${(O.kvCacheBytes/1024/1024/1024).toFixed(2)} GB`);if(console.log(`Total Required Memory: ${((O.modelBytes+O.kvCacheBytes)/1024/1024/1024).toFixed(2)} GB`),O.memoryLimitedCtx!=null){let A=O.memoryLimitedCtx,w=O.kvInfo?.nCtxTrain;if(w)console.log(`
+Memory-Limited Context: ${A} (Train: ${w})`);else console.log(`
+Memory-Limited Context: ${A}`);if(O.limitedKvCacheBytes!=null)console.log(`Limited KV Cache Size: ${(O.limitedKvCacheBytes/1024/1024/1024).toFixed(2)} GB`)}}else if(_.buttress?.selected?.fit){let{totalRequiredBytes:A}=_.buttress.selected.fit;console.log(`Total Required Memory: ${(A/1024/1024/1024).toFixed(2)} GB`)}}if(_.buttress?.selected){let{selected:V}=_.buttress;console.log(`
+=== Hardware Information ===`);let R=null;if(process.platform!=="win32")try{R=Q1("uname -a",{encoding:"utf8"}).trim()}catch{}if(R)console.log(`System: ${R}`);else console.log(`Hostname: ${u.hostname()}`),console.log(`OS: ${u.type()} ${u.release()}`);console.log(`Platform: ${V.platform}`),console.log(`CPU Cores: ${u.cpus().length}`),console.log(`Total System Memory: ${(u.totalmem()/1024/1024/1024).toFixed(2)} GB`);let q=V.cpuTotalBytes>0?(V.cpuUsableBytes/V.cpuTotalBytes*100).toFixed(0):0;if(console.log(`Usable CPU Memory: ${(V.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${q}% of ${(V.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),V.hasGpu)console.log(`
+--- GPU Details ---`),V.devices.filter((w)=>w.type==="gpu").forEach((w)=>{console.log(`GPU Backend: ${w.backend}`),console.log(`GPU Name: ${w.deviceName}`),console.log(`GPU Total Memory: ${(w.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let U=V.gpuTotalBytes>0?(V.gpuUsableBytes/V.gpuTotalBytes*100).toFixed(0):0;if(console.log(`GPU Usable Memory: ${(V.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${U}% of ${(V.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),w.metadata){if(w.metadata.hasBFloat16)console.log("Supports BFloat16: Yes");if(w.metadata.hasUnifiedMemory)console.log("Unified Memory: Yes")}});else console.log("GPU: Not available");if(console.log(`
+Backend Variant: ${V.variant}`),console.log(`Performance Score: ${V.score}`),V.fit){if(console.log(`
+--- Model Fit Analysis ---`),console.log(`Fits in GPU: ${V.fit.fitsInGpu?"Yes":"No"}`),console.log(`Fits in CPU: ${V.fit.fitsInCpu?"Yes":"No"}`),console.log(`Limiting Factor: ${V.fit.limiting}`),V.limitedFit)console.log(`
+--- Memory-Limited Fit Analysis ---`),console.log(`Limited Total Required: ${(V.limitedFit.totalRequiredBytes/1024/1024/1024).toFixed(2)} GB`),console.log(`Fits in GPU (Limited): ${V.limitedFit.fitsInGpu?"Yes":"No"}`),console.log(`Fits in CPU (Limited): ${V.limitedFit.fitsInCpu?"Yes":"No"}`),console.log(`Limiting Factor (Limited): ${V.limitedFit.limiting}`)}}console.log(`
+=== Full Capabilities JSON ===`),console.log(JSON.stringify(_,null,2)),process.exit(0)}catch(X){console.error("Failed to get capabilities:",X.message),process.exit(1)}}async function W4({modelIds:$=[],defaultConfig:Z=null}={}){let X=[];if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.34"}`),console.log(`Generating STT model capabilities comparison...
+`),X.push(`${"@fugood/buttress-server"} v${"2.23.0-beta.34"}`),X.push(`## STT Model Capabilities Comparison
+`),!$||$.length===0)console.error("Error: No model IDs provided"),process.exit(1);try{let j=(U={},Y={})=>{let K=Array.isArray(U)?[...U]:{...U};return Object.entries(Y||{}).forEach(([L,B])=>{if(B&&typeof B==="object"&&!Array.isArray(B))K[L]=j(K[L]||{},B);else K[L]=B}),K},W=Z||{},{server:H,generators:Q=[],...J}=W,N=(U)=>j(JSON.parse(JSON.stringify(J)),U||{}),G=(U)=>{if(Array.isArray(Q)&&Q.length>0){let Y=Q.filter((K)=>K?.type==="ggml-stt");if(Y.length>0&&U){let K=Y.find((L)=>L.model?.repo_id===U);if(K)return N(K)}}return Object.keys(J).length>0?N({}):null},_=[];for(let U=0;U<$.length;U+=1){let Y=$[U],{repoId:K,filename:L}=H3(Y);console.log(`[${U+1}/${$.length}] Analyzing ${Y}...`);let B=G(K);B={...B||{},model:{...B?.model||{},repo_id:K,...L&&{filename:L}}};let E=await r("ggml-stt",null,{config:B,includeBreakdown:!0});_.push({modelId:Y,repoId:K,filename:L,capabilities:E,modelInfo:E.buttress?.selected||null,modelConfig:E.modelConfig||null})}let O=(U)=>U?(U/1024/1024).toFixed(1):"N/A",z=(U)=>U?"✅":"\uD83D\uDEAB";X.push("| Model | Size (MB) | Processing Buffer (MB) | Total Required (MB) | Fits GPU | Fits CPU |"),X.push("|-------|-----------|------------------------|---------------------|----------|----------|"),_.forEach(({modelId:U,modelInfo:Y})=>{let K=O(Y?.modelBytes),L=O(Y?.processingBytes||Y?.kvCacheBytes),B=O(Y?.fit?.totalRequiredBytes),E=z(Y?.fit?.fitsInGpu),M=z(Y?.fit?.fitsInCpu);X.push(`| ${U} | ${K} | ${L} | ${B} | ${E} | ${M} |`)}),X.push(`
+---`),X.push(`
+### System Information`);let V=null;if(process.platform!=="win32")try{V=Q1("uname -a",{encoding:"utf8"}).trim()}catch{}if(V)X.push(`- **System:** ${V}`);else X.push(`- **Hostname:** ${u.hostname()}`),X.push(`- **OS:** ${u.type()} ${u.release()}`);if(X.push(`- **Platform:** ${process.platform}`),X.push(`- **CPU Cores:** ${u.cpus().length}`),X.push(`- **Total System Memory:** ${(u.totalmem()/1024/1024/1024).toFixed(2)} GB`),_.length>0){let Y=_[0].capabilities.buttress?.selected;if(Y){let K=Y.cpuTotalBytes>0?(Y.cpuUsableBytes/Y.cpuTotalBytes*100).toFixed(0):0;if(X.push(`- **Usable CPU Memory:** ${(Y.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${K}% of ${(Y.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),Y.hasGpu){let L=Y.devices.filter((B)=>B.type==="gpu");if(L.length>0){let B=L[0];X.push(`- **GPU Backend:** ${B.backend}`),X.push(`- **GPU Name:** ${B.deviceName}`),X.push(`- **GPU Total Memory:** ${(B.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let E=Y.gpuTotalBytes>0?(Y.gpuUsableBytes/Y.gpuTotalBytes*100).toFixed(0):0;X.push(`- **GPU Usable Memory:** ${(Y.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${E}% of ${(Y.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`)}}else X.push("- **GPU:** Not available")}}X.push(`
+### Command Used`);let R=process.argv.slice(2).join(" ");if(X.push(`\`\`\`bash
+${process.argv[0]} ${process.argv[1]} ${R}
+\`\`\``),X.push(`
+### Package Information`),X.push(`- **Name:** ${"@fugood/buttress-server"}`),X.push(`- **Version:** ${"2.23.0-beta.34"}`),Z&&Object.keys(Z).length>0){X.push(`
+### Configuration`),X.push("<details>"),X.push("<summary>Click to expand TOML configuration</summary>"),X.push("\n```toml");try{let U=N3.stringify(Z);X.push(U)}catch(U){X.push("# Error serializing config"),X.push(JSON.stringify(Z,null,2))}X.push("```"),X.push("</details>")}let A=`ggml-stt-model-capabilities-${new Date().toISOString().replace(/[.:]/g,"-").split("T")[0]}.md`,w=J3.join(process.cwd(),A);Q3.writeFileSync(w,X.join(`
+`),"utf8"),console.log(`
+STT model capabilities table saved to: ${w}`),process.exit(0)}catch(j){console.error("Failed to generate STT model table:",j.message),process.exit(1)}}async function U3({modelId:$=null,defaultConfig:Z=null}={}){if(console.log(`${"@fugood/buttress-server"} v${"2.23.0-beta.34"}`),console.log("Testing capabilities for backend: ggml-stt"),$)console.log(`Model: ${$}`);console.log("--------------------------------");try{let{repoId:X,filename:j}=H3($),W=Z||{},{server:H,generators:Q=[],...J}=W,N=(q={},A={})=>{let w=Array.isArray(q)?[...q]:{...q};return Object.entries(A||{}).forEach(([U,Y])=>{if(Y&&typeof Y==="object"&&!Array.isArray(Y))w[U]=N(w[U]||{},Y);else w[U]=Y}),w},G=(q)=>N(JSON.parse(JSON.stringify(J)),q||{}),O=((q)=>{if(Array.isArray(Q)&&Q.length>0){let A=Q.filter((w)=>w?.type==="ggml-stt");if(A.length>0){if(q){let w=A.find((U)=>U.model?.repo_id===q);if(w)return G(w)}}}if(Object.keys(J).length>0)return G({});return null})(X);if(X)O={...O||{},model:{...J.runtime,...O?.model||{},repo_id:X,...j&&{filename:j}}};let z=await r("ggml-stt",null,{config:O,includeBreakdown:!0}),V=z.buttress?.selected||null,R=z.modelConfig||null;if(X||R?.repoId){console.log(`
+=== Model Information ===`);let q=X||R?.repoId;if(console.log(`Repository ID: ${q}`),j)console.log(`Filename: ${j}`);if(V?.modelBytes)console.log(`Model Size: ${(V.modelBytes/1024/1024).toFixed(1)} MB`);let A=V?.processingBytes||V?.kvCacheBytes;if(A)console.log(`Processing Buffer: ${(A/1024/1024).toFixed(1)} MB`);if(V?.modelBytes&&A)console.log(`Total Required Memory: ${((V.modelBytes+A)/1024/1024).toFixed(1)} MB`);else if(z.buttress?.selected?.fit){let{totalRequiredBytes:w}=z.buttress.selected.fit;console.log(`Total Required Memory: ${(w/1024/1024).toFixed(1)} MB`)}}if(z.buttress?.selected){let{selected:q}=z.buttress;console.log(`
+=== Hardware Information ===`);let A=null;if(process.platform!=="win32")try{A=Q1("uname -a",{encoding:"utf8"}).trim()}catch{}if(A)console.log(`System: ${A}`);else console.log(`Hostname: ${u.hostname()}`),console.log(`OS: ${u.type()} ${u.release()}`);console.log(`Platform: ${q.platform}`),console.log(`CPU Cores: ${u.cpus().length}`),console.log(`Total System Memory: ${(u.totalmem()/1024/1024/1024).toFixed(2)} GB`);let w=q.cpuTotalBytes>0?(q.cpuUsableBytes/q.cpuTotalBytes*100).toFixed(0):0;if(console.log(`Usable CPU Memory: ${(q.cpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${w}% of ${(q.cpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),q.hasGpu)console.log(`
+--- GPU Details ---`),q.devices.filter((Y)=>Y.type==="gpu").forEach((Y)=>{console.log(`GPU Backend: ${Y.backend}`),console.log(`GPU Name: ${Y.deviceName}`),console.log(`GPU Total Memory: ${(Y.maxMemorySize/1024/1024/1024).toFixed(2)} GB`);let K=q.gpuTotalBytes>0?(q.gpuUsableBytes/q.gpuTotalBytes*100).toFixed(0):0;if(console.log(`GPU Usable Memory: ${(q.gpuUsableBytes/1024/1024/1024).toFixed(2)} GB (${K}% of ${(q.gpuTotalBytes/1024/1024/1024).toFixed(2)} GB)`),Y.metadata){if(Y.metadata.hasBFloat16)console.log("Supports BFloat16: Yes");if(Y.metadata.hasUnifiedMemory)console.log("Unified Memory: Yes")}});else console.log("GPU: Not available");if(console.log(`
+Backend Variant: ${q.variant}`),console.log(`Performance Score: ${q.score}`),q.fit)console.log(`
+--- Model Fit Analysis ---`),console.log(`Fits in GPU: ${q.fit.fitsInGpu?"Yes":"No"}`),console.log(`Fits in CPU: ${q.fit.fitsInCpu?"Yes":"No"}`),console.log(`Limiting Factor: ${q.fit.limiting}`)}console.log(`
+=== Full Capabilities JSON ===`),console.log(JSON.stringify(z,null,2)),process.exit(0)}catch(X){console.error("Failed to get capabilities:",X.message),process.exit(1)}}var H3=($)=>{if(!$)return{repoId:null,filename:null};let[Z,X]=$.split(":");return{repoId:Z,filename:X||null}};var G3=S(()=>{$1();B0()});var x$={};import z$ from"node:fs";import B$ from"node:path";import w$ from"@iarna/toml";async function R$($){if(!$?.generators||!Array.isArray($.generators))return;let Z=$.generators.filter((_)=>{if(!_.model?.download)return!1;let{type:O}=_;if(!O||O!=="ggml-llm"&&O!=="ggml-stt")return console.warn(`[Download] Skipping unknown generator type: ${O}`),!1;return!0});if(Z.length===0)return;let{server:X,generators:j,...W}=$,H=Z.map((_)=>{let{type:O}=_,z=_.model?.repo_id;console.log(`[Download] Starting pre-download for ${O}: ${z}`);let V={...W,backend:_.backend||{},model:_.model||{},runtime:{...W.runtime,..._.runtime||{}}};return t0(O,V,{onProgress:()=>{},onComplete:({repoId:R,alreadyExists:q})=>{if(q)console.log(`[Download] Pre-download complete (already exists): ${R}`);else console.log(`[Download] Pre-download complete: ${R}`)},onError:(R)=>{console.error(`[Download] Pre-download failed for ${z}:`,R.message)}})}),Q=await Promise.all(H),J=Q.filter((_)=>_.started).length,N=Q.filter((_)=>_.alreadyExists).length,G=Q.filter((_)=>_.alreadyDownloading).length;console.log(`[Download] Pre-download summary: ${J} started, ${N} already exist, ${G} already downloading`)}var K3,_3,q3,J1,Y0=null,E$,M$,O3,N1,F$;var L3=S(async()=>{G3();S0();X4();await B3();if(process.argv.includes("--version")||process.argv.includes("-v"))console.log("2.23.0-beta.34"),process.exit(0);if(process.argv.includes("--help")||process.argv.includes("-h"))console.log(`
+bricks-buttress v${"2.23.0-beta.34"}
+
+Buttress server for remote inference with GGML backends.
+
+Usage:
+  bricks-buttress [options]
+
+Options:
+  -h, --help                    Show this help message
+  -v, --version                 Show version number
+  -p, --port <port>             Port to listen on (default: 2080)
+  -c, --config <path|toml>      Path to TOML config file or inline TOML string
+
+Testing Options:
+  --test-caps <backend>         Test model capabilities (ggml-llm or ggml-stt)
+  --test-caps-model-id <id>     Model ID to test (used with --test-caps)
+  --test-models <ids>           Comma-separated list of model IDs to test
+  --test-models-default         Test default set of models
+
+  Note: --test-models and --test-models-default output a markdown report
+        file (e.g., ggml-llm-model-capabilities-YYYY-MM-DD.md)
+
+Environment Variables:
+  BUTTRESS_PORT                 Port to listen on (overridden by --port)
+  NODE_ENV                      Set to 'development' for dev mode
+  ENABLE_OPENAI_COMPAT_ENDPOINT Set to '1' to enable OpenAI-compatible API
+
+Examples:
+  bricks-buttress
+  bricks-buttress --port 3000
+  bricks-buttress --config ./config.toml
+  bricks-buttress --test-caps ggml-llm --test-models-default
+  bricks-buttress --test-caps ggml-stt --test-caps-model-id BricksDisplay/whisper-ggml:ggml-small.bin
+`),process.exit(0);K3=process.argv.findIndex(($)=>$==="--port"||$==="-p"),_3=K3>=0?Number(process.argv[K3+1]):void 0,q3=process.argv.findIndex(($)=>$==="--config"||$==="-c"),J1=q3>=0?process.argv[q3+1]:null;if(J1){let $;if(J1.includes(`
+`))$=J1;else{let Z=B$.resolve(J1);try{$=z$.readFileSync(Z,"utf8")}catch(X){console.error(`Failed to read Buttress config at ${Z}:`,X),process.exit(1)}}try{let Z=w$.parse($);if(Z.env&&typeof Z.env==="object")Object.entries(Z.env).forEach(([X,j])=>{if(process.env[X]===void 0)process.env[X]=String(j)}),delete Z.env;Y0=Z}catch(Z){console.error("Failed to parse TOML config:",Z),process.exit(1)}}E$=["ggml-org/gpt-oss-20b-GGUF","ggml-org/gpt-oss-120b-GGUF","unsloth/Nemotron-3-Nano-30B-A3B-GGUF","unsloth/Qwen3-VL-30B-A3B-Instruct-GGUF","bartowski/Mistral-Nemo-Instruct-2407-GGUF","mistralai/Magistral-Small-2509-GGUF","mistralai/Ministral-3-14B-Reasoning-2512-GGUF","bartowski/mistralai_Devstral-Small-2-24B-Instruct-2512-GGUF","bartowski/mistralai_Devstral-2-123B-Instruct-2512-GGUF","ggml-org/gemma-3-12b-it-qat-GGUF","ggml-org/gemma-3-27b-it-qat-GGUF","unsloth/phi-4-GGUF"],M$=["BricksDisplay/whisper-ggml:ggml-small.bin","BricksDisplay/whisper-ggml:ggml-small-q8_0.bin","BricksDisplay/whisper-ggml:ggml-medium.bin","BricksDisplay/whisper-ggml:ggml-medium-q8_0.bin","BricksDisplay/whisper-ggml:ggml-large-v3-turbo.bin","BricksDisplay/whisper-ggml:ggml-large-v3-turbo-q8_0.bin","BricksDisplay/whisper-ggml:ggml-large-v3.bin"],O3=process.argv.findIndex(($)=>$==="--test-caps");if(O3>=0){let $=process.argv[O3+1]||"ggml-llm";if($!=="ggml-llm"&&$!=="ggml-stt")console.error("Only ggml-llm and ggml-stt backends are supported for testing capabilities"),process.exit(1);let Z=process.argv.findIndex((j)=>j==="--test-models"),X=process.argv.includes("--test-models-default");if($==="ggml-stt")if(Z>=0){let j=process.argv[Z+1];if(!j)console.error("Error: --test-models requires a comma-separated list of model IDs"),process.exit(1);let W=j.split(",").map((H)=>H.trim());await W4({modelIds:W,defaultConfig:Y0})}else if(X)await W4({modelIds:M$,defaultConfig:Y0});else{let j=process.argv.findIndex((H)=>H==="--test-caps-model-id"),W=j>=0?process.argv[j+1]:null;await U3({modelId:W,defaultConfig:Y0})}else if(Z>=0){let j=process.argv[Z+1];if(!j)console.error("Error: --test-models requires a comma-separated list of model IDs"),process.exit(1);let W=j.split(",").map((H)=>H.trim());await j4({modelIds:W,defaultConfig:Y0})}else if(X)await j4({modelIds:E$,defaultConfig:Y0});else{let j=process.argv.findIndex((H)=>H==="--test-caps-model-id"),W=j>=0?process.argv[j+1]:null;await V3({modelId:W,defaultConfig:Y0})}}N1=t1(Y0);if(_3)N1.server.port=_3;if(!N1.server.port)N1.server.port=Number(process.env.BUTTRESS_PORT)||2080;F$=process.env.ENABLE_OPENAI_COMPAT_ENDPOINT==="1";z3({config:N1,enableOpenAICompat:F$}).then(async({port:$,openaiEnabled:Z,autoDiscover:X})=>{let j=Y1();if(console.log(`Buttress server listening on port ${$}`),console.log("--------------------------------"),await A3(),console.log(),console.log("Current supported Generators:"),console.log("- LLM (GGML)"),console.log("- STT (GGML)"),console.log(),console.log("Please configure `Buttress (Remote Inference)` in the Generator to connect to this server."),console.log(),console.log(`- Use http://${j}:${$} to connect to this server via LAN.`),console.log(`- Visit http://${j}:${$}/status to see status via LAN.`),console.log(),Z)console.log("OpenAI-compatible API [EXPERIMENTAL]:"),console.log(`- Base URL: http://${j}:${$}/oai-compat/v1`),console.log(`- Chat completions: POST http://${j}:${$}/oai-compat/v1/chat/completions`),console.log(`- Models: GET http://${j}:${$}/oai-compat/v1/models`),console.log();else console.log("OpenAI-compatible API [EXPERIMENTAL]: disabled"),console.log("  Set ENABLE_OPENAI_COMPAT_ENDPOINT=1 to enable"),console.log();if(X)console.log("Auto-discover enabled"),console.log();if(Y0)await R$(Y0)}).catch(($)=>{console.error("Failed to start Buttress server:",$),process.exitCode=1})});import{node as S$}from"@elysiajs/node";import{Elysia as P$,t as P0}from"elysia";import{ReadableStream as T$}from"node:stream/web";import k$ from"node:fs/promises";import{ZodError as D$}from"zod";var b$=async()=>{let $=`https://registry.npmjs.org/${"@fugood/buttress-server"}/latest`;try{let Z=new AbortController,X=setTimeout(()=>Z.abort(),3000),j=await fetch($,{headers:{Accept:"application/json"},signal:Z.signal});if(clearTimeout(X),!j.ok)return null;return(await j.json()).version||null}catch{return null}},v$=($,Z)=>{if(!Z)return!1;let X=$.split(/[.-]/),j=Z.split(/[.-]/);for(let W=0;W<Math.max(X.length,j.length);W+=1){let H=parseInt(X[W])||0,Q=parseInt(j[W])||0;if(Q>H)return!0;if(Q<H)return!1}return!1},h$=($)=>{console.log(""),console.log("\x1B[33m╭─────────────────────────────────────────────────╮\x1B[0m"),console.log("\x1B[33m│\x1B[0m  Update available! \x1B[2m%s\x1B[0m → \x1B[32m%s\x1B[0m","2.23.0-beta.34".padEnd(12),$.padEnd(12),"\x1B[33m│\x1B[0m"),console.log("\x1B[33m│\x1B[0m                                                 \x1B[33m│\x1B[0m"),console.log("\x1B[33m│\x1B[0m  Run to upgrade:                                \x1B[33m│\x1B[0m"),console.log("\x1B[33m│\x1B[0m  \x1B[36mnpm install -g %s\x1B[0m      \x1B[33m│\x1B[0m","@fugood/buttress-server".padEnd(27)),console.log("\x1B[33m╰─────────────────────────────────────────────────╯\x1B[0m"),console.log("")},A3=async()=>{try{let $=await b$();if($&&v$("2.23.0-beta.34",$))h$($)}catch($){}},C$,I$=async({backend:$,router:Z,config:X,enableOpenAICompat:j})=>{try{await k$.mkdir(X.server.temp_file_dir,{recursive:!0})}catch{}let W=Y1()||"0.0.0.0",H={id:X.server.id,name:X.server.name,version:"2.23.0-beta.34",address:W,port:X.server.port,url:`http://${W}:${X.server.port}`,generators:g2(X,X.generators.map((N)=>N.type)),authentication:{required:!0,type:"device-group"}},Q=new P$({serve:{maxRequestBodySize:X.server.max_body_size},websocket:{idleTimeout:Math.ceil(X.server.session_timeout/1000)},adapter:C$?S$():void 0}).state({sessions:new Map,backend:$||e0,config:X,serverInfo:H});if(Z)Q.use(Z);if(X.autodiscover?.http?.enabled)Q.use(n1(X));if(Q.use(s1),Q.use(r1),j)Q.use(X1(X));let J={INVALID_REQUEST:-32600,INVALID_PARAMS:-32602,METHOD_NOT_FOUND:-32601,INTERNAL_ERROR:-32603};return Q.ws("/buttress/rpc",{parse:(N,G)=>{if(typeof G==="string")try{return JSON.parse(G)}catch{return N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.INVALID_REQUEST,message:"Invalid request"},id:null})),null}return G},body:P0.Object({jsonrpc:P0.String(),method:P0.String(),params:P0.String(),id:P0.String()}),open(N){let G=N.id??N.raw?.id??N.remoteAddress;if(console.log(`[Request] New connection: ${G}`),!N.data.store.sessions.has(G))N.data.store.sessions.set(G,{streams:new Map,generators:new Set,timeout:null});else{let _=N.data.store.sessions.get(G);clearTimeout(_.timeout),_.timeout=null}},async message(N,{id:G,method:_,params:O}){let z=N.id??N.raw?.id??N.remoteAddress;console.log(`[Request] Received request from ${z}: ${_}`);let V=N.data.store.sessions.get(z),[R,q]=_.split("."),A=e2[R]?.[q];if(!A){N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.METHOD_NOT_FOUND,message:"Method not found"},id:G}));return}try{if(_==="cancel"){if(V.streams.has(G))V.streams.get(G)?.cancel(),V.streams.delete(G);return}if(_==="ping"){N.send(JSON.stringify({jsonrpc:"2.0",result:"pong",id:G}));return}let w=Z3(O),U=t2[R]?.[q],Y=U?U.parse(w):w,K={...N.data.store,peerId:z,session:V},L=await A(K,...Y);if(L instanceof T$){V.streams.set(G,L),N.send(JSON.stringify({jsonrpc:"2.0",result:{type:"stream"},id:G}));try{let B=L.getReader();while(!0){let{value:E,done:M}=await B.read();if(M)break;let{event:F,data:x}=E;N.send(JSON.stringify({jsonrpc:"2.0",method:`notification/${F}`,params:W1(x),id:G}))}N.send(JSON.stringify({jsonrpc:"2.0",method:"notification/_end",id:G}))}catch(B){console.error(B),N.send(JSON.stringify({jsonrpc:"2.0",method:"notification/_error",params:W1(B),id:G}))}V.streams.delete(G)}else N.send(JSON.stringify({jsonrpc:"2.0",result:W1(L),id:G}))}catch(w){if(w instanceof D$){N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.INVALID_PARAMS,message:"Invalid params",data:w.issues},id:G}));return}console.error(w),N.send(JSON.stringify({jsonrpc:"2.0",error:{code:J.INTERNAL_ERROR,message:String(w)},id:G}))}},async close(N){let G=N.id??N.raw?.id??N.remoteAddress;console.log(`[Request] Connection closed: ${G}`);let{backend:_,sessions:O}=N.data.store,z=O.get(G);if(!z)return;z.streams.forEach((V)=>V.cancel()),z.streams.clear(),z.timeout=setTimeout(()=>{O.delete(G),console.log(`[Request] Session timed out: ${G}`);let{generators:V}=z;V.forEach((R)=>{_.finalizeGenerator(R)})},X.server.session_timeout)}}),{app:Q,config:X}},z3=async({backend:$,router:Z,config:X,enableOpenAICompat:j=!1})=>{let{app:W,config:H}=await I$({backend:$,router:Z,config:X,enableOpenAICompat:j}),{server:{port:Q}}=H,J=[new Promise((G)=>W.listen(Q,G))],N=null;if(H.autodiscover)N=new Z4(H.autodiscover,()=>W.store.serverInfo),J.push(N.start());return await Promise.all(J),{app:W,port:Q,openaiEnabled:j,autoDiscover:N}};var B3=S(async()=>{$1();I2();$3();X3();S0();Y3();X4();$1();S0();C$=typeof process<"u"&&process.versions&&process.versions.node;if(Q4.main==Q4.module)await L3().then(() => x$)});await B3();export{z3 as startServer,t0 as startModelDownload,t1 as processConfig,h$ as logUpdateMessage,I$ as createServer,v$ as compareVersions,b$ as checkForUpdates,A3 as checkAndNotifyUpdates};

package/package.json

@@ -0,0 +1,45 @@
+{
+  "name": "@fugood/buttress-server",
+  "version": "2.23.0-beta.37",
+  "main": "dist/index.mjs",
+  "type": "module",
+  "bin": {
+    "bricks-buttress": "lib/index.mjs"
+  },
+  "files": [
+    "lib",
+    "config"
+  ],
+  "scripts": {
+    "typecheck": "tsc --noEmit",
+    "build": "bun run scripts/build.ts",
+    "dev": "bun src/index.ts",
+    "start": "bun lib/index.mjs",
+    "start-with-node": "node lib/index.mjs"
+  },
+  "keywords": [
+    "BRICKS",
+    "buttress",
+    "server"
+  ],
+  "license": "MIT",
+  "dependencies": {
+    "@elysiajs/cors": "^1.1.1",
+    "@elysiajs/node": "^1.4.2",
+    "@fugood/llama.node": "^1.4.11",
+    "@fugood/whisper.node": "^1.0.11",
+    "@huggingface/gguf": "^0.3.2",
+    "@iarna/toml": "^3.0.0",
+    "bytes": "^3.1.0",
+    "elysia": "^1.4.19",
+    "ms": "^2.1.1",
+    "node-machine-id": "^1.1.12",
+    "zod": "^3.25.76"
+  },
+  "devDependencies": {
+    "bun-plugin-isolated-decl": "^0.2.9",
+    "oxc-transform": "^0.105.0",
+    "typescript": "^5.9.3"
+  },
+  "gitHead": "6c1b834d1a19275da4864fb97e6f24695644c712"
+}